IMAGE RECOGNITION SYSTEM AND METHOD

Abstract

An improved system and method for digital image classification is provided. A host computer having a processor is coupled to a memory storing thereon reference feature data. A graphics processing unit (GPU) having a processor is coupled to the host computer and is configured to obtain, from the host computer, feature data corresponding to the digital image; to access, from the memory, the one or more reference feature data; and to determine a semi-metric distance based on a Poisson-Binomial distribution between the feature data and the one or more reference feature data. The host computer is configured to classify the digital image using the determined semimetric distance.

Claims

1. A computer-implemented method for classifying a digital image, the method comprising: obtaining, from a host computer, feature data corresponding to the digital image; determining, by a graphics processing unit, a semi-metric distance based on a Poisson-Binomial distribution between the feature data and one or more reference feature data stored in a memory of the host computer; and classifying the digital image using the determined semi-metric distance.

2. The method of claim 1, wherein the semi-metric distance is a Poisson-Binomial Radius (PBR).

3. The method of claim 1, wherein classifying the digital image comprises using a support vector machines (SVM) classifier.

4. The method of claim 1, wherein classifying the digital image comprises using a k-nearest neighbor (kNN) classifier.

5. The method of claim 1, wherein the kNN classifier is an adaptive local mean-based k-nearest neighbors (ALMkNN) classifier, wherein the value of the k-nearest neighbors (k) is adaptively chosen.

6. The method of claim 5, wherein the adaptive value of the k-nearest neighbors does not exceed the square root of the number of the one or more reference data.

7. The method of claim 1, wherein the obtained feature data and the one or more reference feature data comprise Pairwise Rotation Invariant Co-occurrence Local Binary Pattern (PRICoLBP) data.

8. The method of claim 1, wherein the obtained feature data and the one or more reference feature data comprise Histogram of Oriented Gradients (HOG) data.

9. The method of claim 1, wherein the obtained feature data comprises an N-dimensional feature vector X such that X=(a.sub.1 . . . a.sub.N) and the reference feature data comprises an N-dimensional feature vector Y such that Y=(b.sub.1 . . . b.sub.N), and wherein the determining the semi-metric distance (PBR(X,Y)) comprises calculating: $\mathrm{PBR}\left(X,Y\right)=\frac{\sigma }{N-\mu }=\frac{\sqrt{\underset{i=1}{\overset{N}{.Math.}}.Math.p_i\left(1-p_i\right)}}{N-\underset{i=1}{\overset{N}{.Math.}}.Math.p_i},$ wherein N is an integer greater than 0, σ is the standard deviation of vector X, μ is the mean of vector X, and p.sub.i is |a.sub.i−b.sub.i|.

10. The method of claim 1, wherein the digital image comprises information corresponding to a DNA or RNA sequence, and the obtained feature data comprises a vector X of sequencing quality proximities for a first DNA sample with a sequencing depth d.sub.x such that X=(x.sub.1 . . . x.sub.dx) and the reference feature data comprises a vector Y of sequencing probabilities for a reference DNA sample with a sequencing depth d.sub.y such that Y=(y.sub.1 . . . y.sub.dy), and wherein the determining the semi-metric distance (PBR.sub.seq) comprises calculating: ${\mathrm{PBR}}_{\mathrm{seq}}\left(X,Y\right)=\frac{{\sigma }_X.Math.{\sigma }_Y}{.Math.{\mu }_X-{\mu }_Y.Math.},$ wherein μ.sub.xis a means for vector X, μ.sub.Y is a means for vector Y, σ.sub.X is a standard deviation for vector X, and σ.sub.Y is a standard deviation for vector Y.

11. The method of claim 10, wherein the classifying the digital image comprises: determining whether the semi-metric distance (PBR.sub.seq) is greater than a threshold value, and classifying the DNA or RNA sequence as being a tumor or normal based on the determining if the semi-metric distance (PBR.sub.seq) is greater than the threshold value.

12. The method of claim 10, wherein classifying the digital image comprises identifying a rare variant in the DNA or RNA sequence.

13. The method of claim 1, further comprising: determining a closest matching reference feature data of the one or more reference feature data.

14. The method of claim 13, further comprising: identifying a person based on the determined closest matching reference feature data, wherein the digital image comprises at least one of: an ear, a face, a fingerprint, and an iris.

15. A system for classifying a digital image comprising: a host computer comprising a processor, wherein the host computer is coupled to a memory comprising one or more reference feature data; and a graphics processing unit (GPU) comprising a processor, wherein the GPU is coupled to the host computer and is configured to: obtain, from the host computer, feature data corresponding to the digital image; access, from the memory, the one or more reference feature data; determine a semi-metric distance based on a Poisson-Binomial distribution between the feature data and the one or more reference feature data; wherein the host computer is configured to: classify the digital image using the determined semi-metric distance.

16. The system of claim 15, wherein the semi-metric distance is a Poisson-Binomial Radius (PBR).

17. The system of claim 15, wherein the host computer is further configured to classify the digital image using a support vector machines (SVM) classifier.

18. The system of claim 15, wherein the host computer is further configured to classify the digital image using a k-nearest neighbor (kNN) classifier.

19. The system of claim 18, wherein the kNN classifier is an adaptive local mean-based k-nearest neighbors (ALMkNN) classifier, wherein the value of the k-nearest neighbors (k) is adaptively chosen.

20. The system of claim 19, wherein the adaptive value of the k-nearest neighbors (k) does not exceed the square root of the number of the one or more reference data.

21. The system of claim 15, wherein the feature data and the one or more reference feature data comprise Pairwise Rotation Invariant Co-occurrence Local Binary Pattern (PRICoLBP) data.

22. The system of claim 15, wherein the obtained feature data and the one or more reference feature data comprise Histogram of Oriented Gradients (HOG) data.

23. The system of claim 15, wherein the feature data comprises an N-dimensional feature vector X such that X=(a.sub.1 . . . a.sub.N) and the reference feature data comprises an N-dimensional feature vector Y such that Y=(b.sub.1 . . . b.sub.N), and wherein the GPU is further configured calculate: $\mathrm{PBR}\left(X,Y\right)=\frac{\sigma }{N-\mu }=\frac{\sqrt{\underset{i=1}{\overset{N}{.Math.}}.Math.p_i\left(1-p_i\right)}}{N-\underset{i=1}{\overset{N}{.Math.}}.Math.p_i},$ wherein PBR(X,Y) is a Poisson-Binomial Radius (PBR) distance between the vector X and the vector Y, N is an integer greater than 0, σ is the standard deviation of vector X, μ is the mean of vector X, and p.sub.i is |a.sub.i−b.sub.i|.

24. The system of claim 15, wherein the digital image comprises information corresponding to a DNA or RNA sequence, and the feature data comprises a vector X of sequencing quality proximities for a first DNA sample with a sequencing depth d.sub.x such that X=(x.sub.1 . . . x.sub.dx) and the reference feature data comprises a vector Y of sequencing probabilities for a reference DNA sample with a sequencing depth d.sub.y such that Y=(y.sub.1 . . . y.sub.dy), and wherein the GPU is further configured to calculate determining the semi-metric distance (PBR.sub.seq) comprises calculating: ${\mathrm{PBR}}_{\mathrm{seq}}\left(X,Y\right)=\frac{{\sigma }_X.Math.{\sigma }_Y}{.Math.{\mu }_X-{\mu }_Y.Math.},$ wherein PBR.sub.seq(X, Y) is a Poisson-Binomial Radius (PBR) distance between the vector X and the vector Y, μ.sub.X is a means for the vector X, μ.sub.Y is a means for the vector Y, σ.sub.X is a standard deviation for the vector X, and σ.sub.Y is a standard deviation for the vector Y.

25. The system of claim 24, wherein the host computer is further configured to: determine whether the semi-metric distance (PBR.sub.seq) is greater than a threshold value, and classify the DNA or RNA sequence as being a tumor or normal based on the determining if the semi-metric distance (PBR.sub.seq) is greater than the threshold value.

26. The system of claim 24, wherein the host computer is further configured to: identify a rare variant in the DNA or RNA sequence.

27. The system of claim 15, wherein the host computer is further configured to: determine a closest matching reference feature data of the one or more reference feature data.

28. The system of claim 27, wherein the host computer is further configured to: identify a person based on the determined closest matching reference feature data, wherein the digital image comprises at least one of: an ear, a face, a fingerprint, and an iris.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0026] FIGS. 1a and 1b illustrate the output probability mass function for DNA sequencing analysis.

[0027] FIGS. 2a and 2b are example images from the (a) LFW dataset and (b) cat dataset, respectively.

[0028] FIG. 3a is a block diagram of exemplary implementation architecture for image recognition according to an embodiment of the present invention.

[0029] FIG. 3b is a block diagram of exemplary implementation architecture for DNA rare variant detection according to an embodiment of the present invention.

[0030] FIG. 3c is a basic flow chart for performing image recognition according to embodiments of the present invention.

[0031] FIG. 4 is a graph of classification accuracy as a function of number of training images.

[0032] FIG. 5 is a bar graph of a comparison of computation time for the image classification application using different distance measures.

[0033] FIGS. 6a and 6b illustrate Cumulative Matching Curves (CMOs) for the (a) IIT Delhi I and (b) IIT Delhi II databases, respectively.

[0034] FIGS. 7a and 7b illustrate the effect of noise on rank-one recognition performance for the (a) IIT Delhi I and (b) IIT Delhi II databases, respectively.

[0035] While the present invention may be embodied in many different forms, a number of illustrative embodiments are described next with reference to the above-described figures, with the understanding that the present disclosure is to be considered as providing examples of the principles of the invention and such examples are not intended to limit the invention to preferred embodiments described herein and/or illustrated herein.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0036] The Poisson-Binomial distribution is defined by the probability mass function for n successes given independent but non-identical probabilities (p.sub.1, . . . , p.sub.N) of success. These events exist in a probability space (Ω,F,P). The distribution is unimodal with mean μ being the summation of p.sub.i where i increments from 1 to N, and variance σ.sup.2 being the summation of (1-p.sub.i)p.sub.i where i increments from 1 to N.

[0037] A special case of this distribution is the binomial distribution where p.sub.i has the same value for all i. The Poisson-Binomial distribution may be used in a broad range of fields such as biology, imaging, data mining, bioinformatics, and engineering. While it is popular to approximate the Poisson-Binomial distribution to the Poisson distribution. This approximation is only valid when the input probabilities are small as evident from the bounds on the error defined by Le Cam's theorem [19] given by

[00002]$\begin{array}{cc}\underset{n=0}{\overset{\infty }{.Math.}}.Math..Math.\left|P\left({\Omega }_n\right)-\frac{{\lambda }_N^n.Math.e^{-{\lambda }_N}}{n!}\right|<2.Math.\underset{i=1}{\overset{N}{.Math.}}.Math..Math.p_i^2.& \left(2\right)\end{array}$

where P(Ω.sub.n) gives the probability of n successes in the Poisson-Binomial domain and λ is the Poisson parameter.

[0038] The Poisson-Binomial distribution has found increasing use in research applications. Shen et al [20] developed a machine learning approach for metabolite identification from large molecular databases such as KEGG and PubChem. The molecular fingerprint vector was treated as Poisson-Binomial distributed and the resulting peak probability was used for candidate retrieval. Similarly, Lai et al. [21] developed a statistical model to predict kinase substrates based on phosphorylation site recognition. Importantly, the probability of observing matches to the consensus sequences was calculated using the Poisson-Binomial distribution. Other groups [22], [23] have used this distribution to identify genomic aberrations in tumor samples.

[0039] Since the probability of an aberration event varies across samples, individual DNA base positions are treated as independent Bernoulli trials with unequal success probabilities for ascertaining the likelihood of a genetic aberration at every position in each sample. Following the same reasoning, models to accurately call rare variants [24], [25] have been proposed.

[0040] The present invention seeks, among other things, to improve the accuracy of DNA sequencing analysis based on sequencing quality scores. Each score made available for every single sequenced DNA base reflects the probability that the output value was correctly called. For example, if there are N independent reads for a particular position, then sequence analysis software will generate a quality score q.sub.i for each read at that position which takes into account the probability of read errors. The implied probability of a correct read is given by

[00003]$\begin{array}{cc}{p}_i=1-{10}^{-\frac{q_i}{10}}.Math..Math.\mathrm{where}.Math..Math.i\in \left[1,N\right]& \left(3\right)\end{array}$

[0041] Because the identity of each sequenced position is called based on multiple reads of the same position, sometimes numbering in the thousands, each read as a Bernoulli event was treated and sought to build a probability distribution for each sequenced position using the relevant quality scores for that position. Efficient ways to compute this probability distribution were found and are described below.

[0042] Using Waring's Theorem

[0043] Define p.sub.1, . . . , p.sub.N as describing independent but non-identical events which exist in a probability space (Ω, F, P). Z.sub.k is furthered defined as the sum of all unique k-combinations drawn from p.sub.1, . . . , p.sub.N. Thus formally:

[00004]$\begin{array}{cc}{Z}_k=\underset{\begin{array}{c}I.Math.\left\{1,\mathrm{.Math.},N\right\}\\ \left|I\right|=k\end{array}}{.Math.}.Math.\underset{i\in I}{.Math.}.Math.p_i,k\in \left\{0,\mathrm{.Math.},N\right\}& \left(4\right)\end{array}$

where the intersection over the empty set is defined as Ω. Thus Z.sub.0=1 and the sum runs over all subsets I of the indices 1, . . . ,N which contain exactly k elements. For example, if N=3, then

Z.sub.1=p.sub.1+p.sub.2+p.sub.3

Z.sub.2=p.sub.1.Math.p.sub.2+p.sub.1.Math.p.sub.3+p.sub.2.Math.p.sub.3

Z.sub.3=p.sub.1.Math.p.sub.2.Math.p.sub.3   (5)

[0044] Then define P(n) in terms of Z.sub.k by normalizing for all redundantly counted set intersections using Waring's theorem [26], which is a special case of the Schuette-Nesbitt formula [27].

[00005]$\begin{array}{cc}P\left({\Omega }_n\right)=\underset{k=n}{\overset{N}{.Math.}}.Math..Math.\left(\begin{array}{c}k\\ n\end{array}\right).Math.{Z_k\left(-1\right)}^{k-n},k\in \left\{0,\mathrm{.Math.},N\right\}.& \left(6\right)\end{array}$

[0045] The inclusion-exclusion theorem is given by n=0. A scalable means of computing Z.sub.k is described in Algorithm 1.

TABLE-US-00001 Algorithm 1 Recursive Waring Algorithm Input P = {p.sub.n ∈   .sup.m}.sub.n=1.sup.N : a probability vector Output: Calculate the values of Z.sub.k given an arbitrary   vector of probabilities  1 Construct an N × N upper triangular matrix A.sup.T =   [a.sub.i,j]  2 a.sub.1,* ← P  3 for i = 2 to N do  4  k ← 0  5  for j = i to N do  6   k ← k + a.sub.i−1,j−1  7   a.sub.i,j ← ka.sub.1,j  8  end for  9  Z.sub.i ← Σ.sub.j=i.sup.N a.sub.i,j 10 end for

[0046] The primary benefit of this approach is the exponentially rising discount in time complexity with increasing values of N. This arises from the dynamic programming character of the algorithm which groups calculations in blocks to minimize redundancy. This self-similar recursive structure makes calculation feasible by avoiding combinatorial explosion. Using this approach, the total number of blocks that need to be calculated increases with N.sup.2 and is described by the arithmetic sum N/2*(1+N).

[0047] Another advantage of this approach is the ability to compute the elements of each column in parallel. This means that the time complexity decreases from O(N.sup.2) without parallelization to O(N) with full parallelization implemented. Further improvement may be made by calculating matrix elements in the reverse direction, thereby providing a tandem method for parallel computation of matrix A.sup.T. This is accomplished by using two recursive functions, a.sub.i,N=a.sub.1,N(Z.sub.i−1−a.sub.i-1,N) and a.sub.i,j=a.sub.i,j.Math.(a.sub.i,j+1/a.sub.1,j+1−a.sub.i−1,j), simultaneously in addition to the recursive function defined in Algorithm 1. The methods described above provide an efficient means of generating joint probability mass functions (p.m.f.). The case for N=6 is demonstrated here with the Z.sub.k series being multiplied by the appropriate binomial coefficients.

[00006]$\left(\begin{array}{ccccccc}\left(\begin{array}{c}0\\ 0\end{array}\right)& -\left(\begin{array}{c}1\\ 1\end{array}\right)& \left(\begin{array}{c}2\\ 2\end{array}\right)& -\left(\begin{array}{c}3\\ 3\end{array}\right)& \left(\begin{array}{c}4\\ 4\end{array}\right)& -\left(\begin{array}{c}5\\ 5\end{array}\right)& \left(\begin{array}{c}6\\ 6\end{array}\right)\\ 0& \left(\begin{array}{c}1\\ 0\end{array}\right)& -\left(\begin{array}{c}2\\ 1\end{array}\right)& \left(\begin{array}{c}3\\ 2\end{array}\right)& -\left(\begin{array}{c}4\\ 3\end{array}\right)& \left(\begin{array}{c}5\\ 4\end{array}\right)& -\left(\begin{array}{c}6\\ 5\end{array}\right)\\ 0& 0& \left(\begin{array}{c}2\\ 0\end{array}\right)& -\left(\begin{array}{c}3\\ 1\end{array}\right)& \left(\begin{array}{c}4\\ 2\end{array}\right)& -\left(\begin{array}{c}5\\ 3\end{array}\right)& \left(\begin{array}{c}6\\ 4\end{array}\right)\\ 0& 0& 0& \left(\begin{array}{c}3\\ 0\end{array}\right)& -\left(\begin{array}{c}4\\ 1\end{array}\right)& \left(\begin{array}{c}5\\ 2\end{array}\right)& -\left(\begin{array}{c}6\\ 3\end{array}\right)\\ 0& 0& 0& 0& \left(\begin{array}{c}4\\ 0\end{array}\right)& -\left(\begin{array}{c}5\\ 1\end{array}\right)& \left(\begin{array}{c}6\\ 3\end{array}\right)\\ 0& 0& 0& 0& 0& \left(\begin{array}{c}5\\ 0\end{array}\right)& -\left(\begin{array}{c}6\\ 1\end{array}\right)\\ 0& 0& 0& 0& 0& 0& \left(\begin{array}{c}6\\ 0\end{array}\right)\end{array}\right).Math.\left(\begin{array}{c}{Z}_0\\ Z_1\\ Z_2\\ Z_3\\ Z_4\\ Z_5\\ Z_6\end{array}\right)=\left(\begin{array}{c}P\left(0\right)\\ P\left(1\right)\\ P\left(2\right)\\ P\left(3\right)\\ P\left(4\right)\\ P\left(5\right)\\ P\left(6\right)\end{array}\right)$

[0048] The same pmf may be generated using an alternative method described below.

[0049] The Fast Fourier Transform

[0050] Using the same definitions as before, the probability of any particular combination ω can be written as the combinatorial product of occurring and non-occurring events.

[00007]$\begin{array}{cc}P\left(\omega \right)=\underset{i\in I}{.Math.}.Math..Math.p_i.Math.\underset{i\in I^C}{.Math.}.Math.\left(1-p_i\right)& \left(7\right)\end{array}$

[0051] If Ω.sub.n is defined to be the corresponding sample space of all possible paired sets of I and I.sup.C resulting from n occurrences and N-n non-occurrences, then

[00008]$\begin{array}{cc}P\left({\Omega }_n\right)=\underset{j=1}{\overset{\left(\begin{array}{c}N\\ n\end{array}\right)}{.Math.}}.Math..Math.\underset{i\in I}{.Math.}.Math..Math.p_i.Math.\underset{i\in I_j^C}{.Math.}.Math.\left(1-p_i\right)& \left(8\right)\end{array}$

The above expression is intuitive as it is the summed probabilities of all possible combinations of occurrences and non-occurrences. By observation, it is possible to construct a polynomial to express P(Ω.sub.n) as coefficients of an N-order polynomial.

[00009]$\begin{array}{cc}\underset{i=1}{\overset{N}{.Math.}}.Math..Math.\left[p_i.Math.x+\left(1-p_i\right)\right]=\underset{n=0}{\overset{N}{.Math.}}.Math..Math.P\left({\Omega }_n\right).Math.x^n& \left(9\right)\end{array}$

[0052] The coefficients for the above polynomial may then by easily solved using algorithms based on the Discrete Fourier Transform. The relevant coefficient vector may be efficiently calculated as follows:

[00010]$\begin{array}{cc}C\left(x\right)=\mathrm{IFFT}\left(\underset{k=1}{\overset{N}{.Math.}}.Math.\mathrm{FFT}\left(\begin{array}{cc}{p}_k& 1-p_k\end{array}\right)\right)& \left(10\right)\end{array}$

[0053] Practically speaking, the vectors may be padded with leading zeros to a power of two length and then iteratively processed in pairs using IFFT.sup.−1(FFT(a).Math.FFT(b)) where a and b represent any arbitrary pair of vectors. Using a GPU-implementation of the Fast Fourier Transform (FFT), multiple inputs can be processed in parallel using a simple scheme of interleaved inputs and de-convoluted outputs. This function returns a list of tuples, where the i.sup.th tuple contains the i.sup.th element from each of the argument sequences or iterables.

[0054] DNA Sequencing.

[0055] One important application of the present invention is the analysis of next-generation DNA sequencing datasets where thousands of reads have to be analyzed per DNA base position. If a particular base position is mutated in cancer, then detection of such variants would be an ideal diagnostic. In reality, variant DNA is often admixed with normal DNA in low proportions and the challenge is to calculate statistical confidence given two conflicting states detected at the same base position. This may be accomplished by treating these conflicting states as Bernoulli events and constructing p.m.f.s using either of the two methods described above. Example output is illustrated in FIGS. 1a and 1b.

[0056] Confidence intervals calculated from these p.m.f.s then allow a decision as to whether evidence for the variant base state is sufficiently above a threshold of significance. According to aspects of the present invention, similar principles can be applied to pattern recognition applications, especially those pertaining to image analysis. This can be supported by the fact that pixel intensity can only be considered as a random variable and it does not have a true value, since it is governed by the laws of quantum physics [28].

[0057] Poisson-Binomial Radius Semimetric Distance

[0058] Calculating confidence intervals for every pairwise distance comparison would be computationally intensive in a large image dataset. To avoid this cost and improve efficiency, a distance measure can be defined for independent but non-identically distributed feature descriptors as follows:

[0059] Definition. Given two N dimensional feature vectors X=(a.sub.1, a.sub.2, a.sub.3, . . . , a.sub.N) and Y=(b.sub.1, b.sub.2, b.sub.3, . . . , b.sub.N) with p.sub.i=|a.sub.i−b.sub.i|, the distance between the two vectors is

PB.sub.m(X, Y)=P(Ω.sub.m) (N−m)   (11)

where m is the mode and P(m) is the peak probability of the distribution. Darroch [29] has previously shown that mode m may be bounded as follows:

[00011]$m=\{\begin{array}{cc}n,& \mathrm{if}.Math..Math.n\le \mu \le n+\frac{1}{n+2}\\ n.Math..Math.\mathrm{and}.Math.\text{/}.Math.\mathrm{or}.Math..Math.n+1,& \mathrm{if}.Math..Math.n+\frac{1}{n+2}\le \mu \le n+1-\frac{1}{N-n+1}\\ n+1,& \mathrm{if}.Math..Math.n+1-\frac{1}{N-n+1}<\mu \le n+1\end{array}$

where 0≦n≦N. This implies that m differs from the mean μ by less than 1. Thus, although the mode m is a local maximum, it is approximated by the mean μ. This allows

PB.sub.μ(X, Y)=P(Ω.sub.μ) (N−μ)   (12)

[0060] A further refinement may be made by considering the excess kurtosis of the Poisson-Binomial distribution which is given by

[00012]$\begin{array}{cc}\mathrm{Kurt}\left[\underset{n=0}{\overset{N}{.Math.}}.Math..Math.P\left({\Omega }_n\right)\right]=\frac{1}{{\sigma }^2}-6& \left(13\right)\end{array}$

where σ.sup.2 is the variance of the p.m.f. The inverse relationship between the peakness of the distribution with σ.sup.2 implies a similar relationship between P(Ω.sub.μ) and σ. This inverse relationship is also consistent with the work of Baillon et al [30] which established the following sharp uniform upper bound for sums of Bernoulli trials.

[00013]$\begin{array}{cc}P\left({\Omega }_n\right)\le \frac{\eta }{\sigma }& \left(14\right)\end{array}$

where η is the upper bound constant. The implication of this inverse relationship is that σ can be adopted as a surrogate measure of P(Ω.sub.μ), thereby avoiding the need to generate a p.m.f. for each distance calculation. Thus, the following semi-metric for independent and non-identical feature descriptors can be defined.

[0061] Given two N dimensional feature vectors X=(a.sub.1, a.sub.2, a.sub.3, . . . , a.sub.N) and Y=(b.sub.1, b.sub.2, b.sub.3, . . . , b.sub.N) with p.sub.i=|a.sub.i−b.sub.i|, the Poisson-Binomial Radius distance between the two vectors is

[00014]$\begin{array}{cc}\mathrm{PBR}\left(X,Y\right)=\frac{\sigma }{N-\mu }=\frac{\sqrt{\underset{i=1}{\overset{N}{.Math.}}.Math.p_i\left(1-p_i\right)}}{N-\underset{i=1}{\overset{N}{.Math.}}.Math.p_i}& \left(15\right)\end{array}$

PBR(X, Y) is a semi-metric. A function d: X×X.fwdarw.[0, 1] is semi-metric over a set X if it satisfies the following properties for {x, y} X: (1) Non-negativity, d(x,y)>=0; (2) Symmetry property, d(x,y)=d(y,x); and 3) Reflexivity, d(x,x)=0. PBR is a nonnegative function and satisfies the reflexivity property. Because only absolute values are used, PBR also satisfies the symmetry property. See Table 4 below showing that PBR and PB.sub.μ are equivalent distance measures for practical purposes.

[0062] Image Classification Application

[0063] Image classification is a computer automated process of assigning a digital image to a designated class based on an analysis of the digital content of the image (e.g., analysis of pixel data). The most common use of such processes is in image retrieval, or more specifically, content-based image retrieval (CBIR). CBIR is the process of retrieving closely matched or similar images from one or more digital image repositories based on automatically extracted features from the query image. It has found numerous practical and useful applications in medical diagnosis, intellectual property, criminal investigation, remote sensing systems and picture archiving and management systems. See [31].

[0064] The key objectives in any CBIR system are high retrieval accuracy and low computational complexity (the present invention improves both). Implementing an image classification step prior to image retrieval can increase retrieval accuracy. Moreover, the computational complexity can also be reduced by this step.

[0065] Let N.sub.T be the number of training images per class, N.sub.C the number of classes and N.sub.D the number of feature descriptors per image. The computational complexity of a typical CBIR system is O(N.sub.T.Math.N.sub.C.Math.N.sub.D+(N.sub.T.Math.N.sub.C).Math.log(N.sub.T.Math.N.sub.C)). See [34]. In contrast adding a pre-classification step decreases complexity to O(N.sub.C.Math.N.sub.D.Math.log(N.sub.T.Math.N.sub.D))+O(N.sub.T.Math.N.sub.D+N.sub.T.Math.log(N.sub.T)). The first term refers to image pre-classification using a Naive-Bayes nearest neighbor classifier [35] and the second term refers to the CBIR process itself.

[0066] To give some perspective, consider the case of N.sub.T=100, N.sub.C=10 and N.sub.D=150. The computational complexity of the latter compared to the former results in an increase in processing speed of 7 times. Thus, image pre-classification improves CBIR performance.

[0067] The detection of cat heads and faces have attracted the recent interest of researchers, reflecting their popularity on the internet and as human companions [36], [37], [38], [39]. Cats present interesting challenges to pattern recognition. Although, sharing a similar face geometry to humans, approaches for detecting human faces can't be directly applied to cats because of the high intra-class variation among the facial features and textures of cats as compared to humans. The present invention is a PBR-based classifier that can distinguish between the two.

[0068] The Labelled Faces in the Wild (LFW) image dataset (FIG. 2a) was created by the authors of [40], and the cat dataset (FIG. 2b) was created by the authors of [36]. These consist of 13,233 human images and 9,997 cat images. In an example, within each class, 70% of the images were randomly partitioned for training and the remaining 30% for testing.

[0069] According to aspects of the present invention, an image classification system is shown in FIG. 3a that is capable of performing image classification as described herein (see, also, basic process shown in FIG. 3c). As shown, the system includes a graphics processor coupled with a host computing system (e.g., CPU), that has access to memory (not shown). As shown in FIG. 3a, the Host is capable of accessing stored images or image data for training images. As described below, each image is resized to a standard size (e.g., 250×250 pixels) which can be selected based on the particular application. After resizing, the Histogram of Oriented Gradients (HOG) is used for feature extraction. The HOG data may be stored and accessed for each training image so that it need not be created (or recreated) on the fly. The image to be classified (on FIG. 3a, the test image) is received at the Host from an image source such as memory, a network, or an image capture system (camera, scanner, or other imaging device). The image is resized to the standard size. After resizing, the Histogram of Oriented Gradients (HOG) is used for feature extraction.

[0070] The HOG data is input into the GPU for further processing. Orientation is computed and a histogram created. The histogram is normalized (as shown, by the Host). PBR calculation is performed on both the training image data and the test image data. Of course, the PBR calculation may be performed ahead of time for the training images and the results stored. Finally, comparisons are made to classify the image by finding the closest match using the PBR results. For example, algorithm 2 (below) may be employed.

[0071] In one example, a GPU-parallelized version of Histogram of Oriented Gradients (HOG) [41] was used for feature extraction. A classifier called adaptive local mean-based k-nearest neighbor (ALMKNN) is used, which is a modification of a local mean-based nonparametric classifier used in Mitani et al [42]. ALMKNN is partially implemented on the GPU.

[0072] HOG features may be GPU-implemented using NVIDIAs Compute Unified Device Architecture (CUDA) framework. HOG features were first described by Navneet Dalai and Bill Triggs [41] as a means of abstracting appearance and shape by representing the spatial distribution of gradients in an image. This has been applied in pedestrian detection [43], vehicle detection [44] and gesture recognition [45]. According to one embodiment, the rectangular-HOG (R-HOG) variant [46] is used as described below.

[0073] According to another aspect of the present invention, a system and method for DNA sequencing, such as rare variant detection, for example, in the case of a tumor biopsy, may be provided. A vector X of sequencing quality probabilities is from an input DNA sample at a single base position with sequencing depth d.sub.x such that X=(x.sub.1, x.sub.2, x.sub.3, . . . . , x.sub.dx), and a similar vector Y from a reference DNA sample sequenced to depth d.sub.y such that Y=(y.sub.1, y.sub.2, y.sub.3, . . . , y.sub.dy). Calculate the means (μ) and standard deviations (σ) for both vectors as follows

[00015]$\begin{array}{cc}{\mu }_Y=\frac{1}{d_y}.Math.{.Math.}_{i=1}^{d_y}.Math..Math.y_i& {\mu }_X=\frac{1}{d_x}.Math.{.Math.}_{i=1}^{d_x}.Math..Math.x_i\\ {\sigma }_Y=\sqrt{{.Math.}_{i=1}^{d_y}.Math.\left(1-y_i\right).Math.y_i}& {\sigma }_X=\sqrt{{.Math.}_{i=1}^{d_x}.Math.\left(1-x_i\right).Math.x_i}\end{array}$

[0074] To compare vectors X and Y, PBR.sub.seq can be defined to be the distance between X and Y as follows:

[00016]${\mathrm{PBR}}_{\mathrm{seq}}\left(X,Y\right)=\frac{{\sigma }_X.Math.{\sigma }_Y}{.Math.{\mu }_X-{\mu }_Y.Math.}$

[0075] A small PBR.sub.seq value indicates a greater likelihood of a tumor sample. For the purposes of classification, a simple threshold T can be defined such that sample X is classified as a tumor if PBR.sub.seq≦T but otherwise classified normal.

[0076] As shown in FIG. 3b, a system for DNA sequencing is provided. As shown in FIG. 3b, the system is similar to the system of 3a, but implements the above method for DNA sequencing using vector data. As shown, the input quality score vector as described above is converted to an input probability vector, which can be accomplished by the host. Reference probability vectors can be provided ahead of time or calculated by the host and provided to the GPU. The GPU is configured to receive the two probability vectors and to compute the PBR.sub.seq distance between the input and reference vectors. The distance is used to classify the DNA sequence and the host outputs an indication of the assigned class.

[0077] Gradient Computation.

[0078] Given an input image 1(x, y), 1-D spatial derivatives I.sub.x(x, y) and I.sub.y(x, y) may be computed by applying gradient filters in x and y directions. The gradient magnitude Mag(x, y) and orientation (x, y) for each pixel may be calculated using:

Mag(x, y)=√{square root over (I.sub.x(x, y).sup.2+I.sub.y(x, y).sup.2 )}  (16)

θ(x, y)=tan.sup.−1(I.sub.y(x, y)/I.sub.x(x, y))   (17)

[0079] Histogram Accumulation.

[0080] The histograms may be generated by accumulating the gradient magnitude of each pixel into the corresponding orientation bins over local spatial regions called cells. In order to reduce the effect of illumination and contrast, the histograms are normalized across the entire image. Finally, the HOG descriptor is formed by concatenating the normalized histograms of all cells into a single vector.

[0081] In one example, the HOG algorithm described above was implemented using the PyCUDA toolkit [47] version 2012.1 and version 5.0 of the NVIDIA CUDA Toolkit and executed on a GeForce GTX 560 Ti graphics card. Each image was resized to 250×250 (62,500 pixels) then subdivided equally into 25 cells, each 50×50 pixels. To address 62,500 pixels, 65,536 threads are created in the GPU having 32×32 threads per block and 8×8 blocks per grid. After allocating memory in both the host and GPU, the kernel is launched.

[0082] Gradient magnitudes, orientations and the histogram may be calculated after the histogram is transferred to the Host where normalization across the entire image is carried out.

[0083] Classification Module

[0084] Classifiers may be either parametric or non-parametric. Parametric classifiers assume a statistical distribution for each class which is generally the normal distribution. Training data is used only to construct the classification model and then completely discarded. Hence they are referred to as model-based classifiers or eager classifiers. In comparison, non-parametric classifiers make no assumption about the probability distribution of the data, classifying the testing tuple based only on stored training data and are therefore also known as instance-based or lazy classifiers. The prototypical example of a parametric classifier is the support vector machine (SVM) algorithm which requires an intensive training phase of the classifier parameters [48], [49], [50] and conversely, one of the most well-known non-parametric classifiers is the k-nearest neighbor (kNN) classifier.

[0085] kNN [51] has been widely used in pattern recognition problems due to its simplicity and effectiveness. Moreover, it is considered to be one of the top ten algorithms in data mining [52]. kNN assigns each query pattern a class associated with the majority class label of its k-nearest neighbors in the training set. In binary (two-class) classification problems, the value of k is usually an odd number to avoid tied votes. Even though kNN has several advantages such as the ability to handle a huge number of classes, avoidance of over fitting and the absence of a training phase, it suffers from three major drawbacks: (1) computational time, (2) the influence of outliers [53] and (3) the need to choose k [54].

[0086] The first problem, time complexity, arises during the computation of distances between the training set and query pattern, particularly when the size of the training set is very large. This problem can be addressed by parallelizing kNN, reducing time complexity to a constant O(1). This compares well to alternative implementations such as search trees which are O(log N) in time. The second problem involves the influence of outliers. To get around this problem, an approach focusing on local neighbors can be used. This type of approach, called LMKNN (Local Mean kNN) however, still carries the problem of having to choose a value for k. Most of the time, k is chosen by cross validation techniques [55].

[0087] This is, however, time consuming and carries the risk of over fitting. Thus, the present invention involves an algorithm where k is adaptively chosen, thus obviating the need for a fixed k value. To put an upper bound on k, the general rule of thumb is used, which is the square root of N, where N is the total training instances in T [56]. The algorithm is referenced as the Adaptive LMKNN or ALMKNN.

[0088] The workings of this classifier are described in Algorithm 2.

TABLE-US-00002 Algorithm 2 ALMKNN Algorithm Input: Query pattern: x ; T = {x.sub.n ∈   .sup.m }.sub.n=1.sup.N : a TS; c.sub.1, c.sub.2, ...,   c.sub.M: M class labels ; k.sub.min: Minimum number of nearest neighbours;   k.sub.max: Maximum number   of nearest neighbours; LB: Lower bound; UB: Upper bound Output: Assign the class label of a query pattern to a nearest local mean vector among classes  1 Compute the distances between x to all x.sub.i belonging to T  2 Choose k.sub.min nearest neighbour in T, say T.sub.k.sub.min(x)  3 Determine the number of neighbours in T.sub.k.sub.min(x) for each class c.sub.i  4 if all the members in T.sub.k.sub.min(x) represents only one class c then  5  Assign x to class c  6 else  7  while k.sub.min ≦ k.sub.max do  8   count ← 0  9   Calculate the local mean vector for each class c.sub.i in set T.sub.k.sub.min(x) 10   Calculate the distances d.sub.i between x and each local mean vector 11   Sort d.sub.i in ascending order to get d.sub.1 < d.sub.2 < d.sub.3.... < d.sub.M 12   Calculate the percentage change between the nearest two distances     d.sub.1 and d.sub.2 13   if percentage change > LB + ((UB − LB)/(k.sub.max − k.sub.min)) then 14    Assign x to class c[d.sub.1] 15    break while 16   else 17    k.sub.min ← k.sub.min + 1 18    count ← count + 1 19   end if 20  end while 21  else  22  Calculate local mean vector for each class c.sub.i in set T.sub.k.sub.min(x)  23  Calculate the distances d.sub.i between x and each local mean vector  24  Assign x to the class c.sub.i with a nearest local mean vector 25 end if

[0089] With 16,261 (N in T) training instances, the limits on neighbors, k.sub.min and k.sub.max, may be defined as 20 and 127 (floor of √N) respectively. A lower bound (LB) and upper bound (UB) may be defined for decision-making to be 2% and 50% respectively. The first step of distance computation maybe implemented in GPU using CUDAMat [57]. The rest of the algorithm was implemented in CPU (Host). There is no training phase and HOG descriptors for the training images are stored in memory.

[0090] Classification Performance

[0091] Using ALMKNN as the framework, the various distance measures were evaluated side-by-side, namely PBR, L.sub.0.1, L.sub.0.5, L.sub.1 and L.sub.2. Classification accuracy of the present invention was averaged over six runs of repeated random sub-sampling validation and these results are shown in FIG. 4. Interestingly, PBR and L.sub.1 were nearly identical in accuracy, easily outperforming the other distance measures. Euclidean distance was able to perform marginally better with a small training set but quickly lost out as the number of training images increased.

[0092] Effect of Noise:

[0093] To test if PBR would be more resistant to noise degradation compared to other distance measures, both training and testing images were corrupted with salt and pepper noise of increasing density, d. At d=0, PBR significantly outperformed all distance measures except L.sub.1. However, consistent with our hypothesis, PBR significantly outperformed all distance measures including L.sub.1 when even a minimal amount of noise (d=0.05) was added (Table 1).

TABLE-US-00003 TABLE 1 A comparison of AUC achieved by the 5 methods. Noise level (d) L.sub.0.1 L.sub.0.5 L.sub.1 L.sub.2 PBR 0.00 0.9758* 0.9899* 0.9916 0.9892* 0.9918 0.05 0.9061* 0.9246* 0.9295* 0.9224* 0.9336 0.25 0.8348* 0.8543* 0.8572* 0.8443* 0.8607 0.50 0.7938* 0.8229* 0.8326* 0.8292* 0.8364

[0094] The Area Under the Curve (AUC) for each method was averaged over 6 independent runs of repeated random sub-sampling validation. The Wilcoxon signed-rank test with 95% confidence level was used to compare other methods with PBR. Methods which performed significantly worse than PBR are highlighted with an asterisk *. The highest AUC for each noise level is bolded.

[0095] Computation Time

[0096] Computation time was measured on a 64-bit Intel Core i5-3470 CPU @ 3.20 GHz 12 GB RAM PC system running Ubuntu 12.04 LTS.

TABLE-US-00004 TABLE 2 Average computation time for processing a 250 × 250 pixel image. CPU (ms) GPU (ms) Speedup Grab Image 0.007 — — HoG 17.19  5.73 3   PBR Calculation 31.7 12.11 2.6 Classifier 1.37 — — Total 50.26 19.21 2.6
From Table 2, it can be seen that a GPU implementation of the present invention was roughly 2.6 times faster than a purely CPU version. This speed up brings PBR almost on par with L.sub.1 and L.sub.2. Computation time was further reduced by introducing the Nearest Mean Classifier (NMC) (Algorithm 3) as a step before the ALMKNN classifier. A confidence measure (CM) of 20% was used, meaning that the NMC result was used for classification when the contrast between distances to the centroids exceeded 20%.

TABLE-US-00005 Algorithm 3 The Nearest Centroid and ALMKNN Algorithm Input: Query pattern: x ; T = {x.sub.n ∈   .sup.m}.sub.n=1.sup.N : a TS; c.sub.1, c.sub.2, ...,   c.sub.M: M class labels ; k.sub.min: Minimum number of nearest neighbours;   k.sub.max: Maximum number of nearest neighbours; LB: Lower bound;   UB: Upper bound; CM: Confidence measure Output: Assign the class label of a query pattern to a nearest local mean vector among classes  1 Calculate the local mean vector for each class c.sub.i belonging to T  2 Calculate the distances d.sub.i between x and each local mean vector using   normalised Manhattan distance measure  3 Sort d.sub.i in ascending order to get d.sub.1 < d.sub.2 < d.sub.3.... < d.sub.M  4 Calculate the percentage change between the nearst two distances   d.sub.1 and d.sub.2  5 if percentage change > CM then  6  Assign x to class c[d.sub.1]  7 else  8  Assign x to class c.sub.i using ALMKNN classifier  9 end if
Accuracy results were exactly the same but computation time was significantly improved, as shown in FIG. 5.

Ear Biometrics Application

[0097] Biometric technology deals with automated methods of verifying the identity of an individual using traits which may be physiological or behavioral. The field of automated biometrics has made significant advances over the last decade with face, fingerprints and iris biometrics having emerged as the most commonly implemented modalities. No single biometric modality is free of shortcomings. Face biometrics for example, has been extensively researched and but yet is prone to failure in sub-optimal conditions [58], [59].

[0098] While fingerprints are complex enough in theory to give a unique signature, in reality, fingerprint biometrics are not spoof-proof, since the system is vulnerable to attack by fake fingerprints made of gelatin, silicon and latex [60]. Iris biometrics has proven to be highly accurate and reliable but its performance deteriorates rapidly under poor lighting, target movement, aging, partial occlusion by eyelids and sensitivity to acquisition instance. This has motivated research into other traits which can overcome the problems of the more established biometrics. One of these new traits, ear biometrics has received increasing attention for a host of reasons. [0099] 1) Unlike faces and irises, ear shape is reasonably invariant over teenage and adult life. Any changes generally occur before the age of 8 and after 70 [61]. [0100] 2) A controlled environment is not required for ear imaging because image context takes its reference from the side of face. [0101] 3) Ear biometrics is able to distinguish between genetically identical twins whereas face biometrics fails in this respect [62]. [0102] 4) The ear has a more uniform distribution of color and less variability with facial expressions.

[0103] According to aspects of the present invention, an ear recognition system is provided based on HOG features and PBR, in accordance with the description above. Databases that have been used are IIT Delhi Ear Databases I and II [63]. There are 125 subjects and 493 images in IIT Delhi DB 1; 221 subjects and 793 images in IIT Delhi DB 2.

[0104] The testing image for each subject in both the databases was randomly picked and the remaining images were used for training.

Biometric Analysis Architecture

[0105] There are three main steps in an ear recognition system: (1) pre-processing (2) feature extraction and (3) template matching. Histogram equalization maybe used as a pre-processing step. Feature extraction may be as already described above. According to aspects of the present invention, a matching module may search for the closest match among the training images. The images in these databases were 50×180 pixels and were resized to 50×50 pixels.

Recognition Performance

[0106] Performance was evaluated using rank-one recognition accuracy. The recognition result was averaged over ten runs. The mean and standard deviation of rank-one recognition rate for all the distance measures are shown in Table 3.

TABLE-US-00006 TABLE 3 Rank-one recognition performance on IIT Delhi databases IIT Delhi I IIT Delhi II PBR 92.9 ± 1.7 93.0 ± 1.2 L.sub.0.1 90.7 ± 2.4 90.3 ± 1.3 L.sub.0.5 92.7 ± 1.9 93.0 ± 1.1 L.sub.1 92.6 ± 1.7 92.9 ± 1.2 L.sub.2 89.3 ± 1.6 89.8 ± 1.5

[0107] Cumulative Match Curves (CMCs) are used to measure performance for biometric recognition systems and have been shown to be directly related to the Receiver Operating Characteristic curve (ROC) in the context of performance verification[64]. Hence, also shown are the CMCs for all the measures in FIG. 6.

Effect of Noise:

[0108] In an experiment, the present invention was applied to training and test images that were corrupted with salt and pepper noise of increasing density, d. Comparisons are shown in FIGS. 7a and 7b. It can be seen that all distance measures except L.sub.2 are stable over noise with L.sub.2 performance degrading sharply with increasing noise density d.

Correlation Between PBμ and the Distance Measures:

[0109] Taking the rank ordering of images matched by the various distance measures to a defined test image, the correlation between PBμ and the other measures (i.e. PBR, L.sub.0.1, L.sub.0.5, L.sub.1 and L.sub.2) was taken. The results in Table 4 show that PBR and PB.sub.μ are highly correlated and rank order is virtually identical between these two distance measures. This is consistent with PB.sub.μ and PBR being approximately equivalent distance measures.

TABLE-US-00007 TABLE 4 Spearman rank correlation coefficient between PB.sub.μ and other distance measures for one testing image PB.sub.μ PBR 0.9995 L.sub.0.1 0.8452 L.sub.0.5 0.9887 L.sub.1 0.9889 L.sub.2 0.9738

Kernel-Based Image Classification

[0110] PBR is a distance metric accept different inputs (PRICoLBP, HOG) and also work within different machine learning frameworks (KNN, SVM kernel).

[0111] Although SVMs (Support Vector Machines) require input data to be independent and identically distributed, they are applied successfully in non-i.i.d scenarios such as speech recognition, system diagnosis etc. [65]. Hence, SVM framework may be employed to illustrate the efficiency of PBR distance in image classification. To incorporate PBR into the SVM framework, the following generalized form of RBF kernels is used [66]:

K.sub.d−RBF(X, Y)=e.sup.−pd(X,Y)

[0112] Where ρ is a scaling parameter obtained using cross-validation and d(X, Y) is the distance between two histograms X and Y. Distance may be defined using a slightly modified form of PBR as follows:

[0113] Definition. Given two N dimensional feature vectors X=(a.sub.1, a.sub.2, a.sub.3, . . . , a.sub.N) and Y=(b.sub.1, b.sub.2, b.sub.3, . . . , b.sub.N) with p.sub.i=a.sub.i ln(2a.sub.i/(a.sub.i+b.sub.i))+b.sub.i ln(2 b.sub.i/(a.sub.i+b.sub.i)), the distance between the two vectors is:

[00017]$d\left(X,Y\right)=\frac{\underset{i=1}{\overset{N}{.Math.}}.Math.p_i\left(1-p_i\right)}{N-\underset{i=1}{\overset{N}{.Math.}}.Math.p_i}.$

[0114] The PBR kernel may be obtained by substituting d(X, Y) into the SVM framework.

Experiments

[0115] The performance of the PBR distance kernel was evaluated in the following six different applications: texture classification, scene classification, species, material, leaf, and object recognition. The texture data sets are Brodatz [67], KTH-TIPS [68], UMD [69] and Kylberg [70]. The scene classification application was based on the Scene-15 [71] data set. For recognition tasks, the Leeds Butterfly [72], FMD [73], Swedish Leaf [74] and Caltech-101 [75] data sets were employed. For both classification and recognition tasks, the dependence of performance on the number of training images per class was evaluated. In each data set, n training images were randomly selected, and the remaining for testing, except in the Caltech-101 data set where the number of test images was limited to 50 per class. All experiments were repeated a hundred times for texture data sets and ten times for others. For each run, the average accuracy per category was calculated. This result from the individual runs was used to report the mean and standard deviation as the final results. Only the grayscale intensity values for all data sets was used, even when color images were available.

[0116] Multi-class classification was done using the one-vs-the-rest technique. For each data set, the SVM hyper-parameters such as C and gamma was chosen by cross-validation in the training set with

C ∈ [2.sup.−2, 2.sup.18]

and

gamma ∈ [2.sup.−4, 2.sup.10] (step size 2).

[0117] Recently, the Pairwise Rotation Invariant Co-occurrence Local Binary Pattern (PRICoLBP) feature has been shown to be efficient and effective in a variety of applications [76]. The significant attributes of this feature are rotational invariance and effective capture of spatial context co-occurrence information. Hence, this feature was used for experiments.

Texture Classification

[0118] The Brodatz album is a popular benchmark texture data set which contains 111 different texture classes. Each class comprises one image divided into nine non-overlapping sub-images.

[0119] The KTH-TIPS data set consists of 10 texture classes, with 81 images per class. These images demonstrate high intra-class variability since they are captured at nine scales under three different illumination directions and with three different poses.

[0120] The UMD texture data set contains 25 categories with 40 samples per class. These uncalibrated, unregistered images are captured under significant viewpoint and scale changes along with significant contrast differences.

[0121] The Kylberg data set has 28 texture classes of 160 unique samples per class. The classes are homogeneous in terms of scale, illumination and directionality. The ‘without’ rotated texture patches version of the data set were used.

[0122] The 2.sub.a template configuration of PRICoLBP was used, which yielded 1,180 dimensional feature for all data sets. Experimental results are shown in Tables 5, 6, 7, and 8 for Brodatz, KTH-TIPS, UMD and Kylberg data sets respectively. From the results, we observe that PBR consistently outperforms other methods when the number of training images is low and yields smaller standard deviations when compared to other distance measures along with a higher classification rate.

TABLE-US-00008 TABLE 5 Texture Classification Results (Percent) on Brodat Training images per class Methods 2 3 PBR 95.9 ± 0.6 96.8 ± 0.5 BD 95.6 ± 0.9 96.6 ± 0.5 JD 95.7 ± 0.7 96.6 ± 0.6 χ.sup.2 95.5 ± 1.0 96.6 ± 0.5 L.sub.1 93.1 ± 1.1 94.9 ± 0.8 L.sub.2 89.2 ± 1.4 92.3 ± 1.0 L.sub.0.5 92.9 ± 0.8 94.4 ± 0.8 L.sub.1-BRD 93.1 ± 1.1 95.0 ± 0.9 HI 93.1 ± 1.0 95.0 ± 0.8 Hellinger 94.2 ± 0.9 95.8 ± 0.6

TABLE-US-00009 TABLE 6 TEXTURE CLASSIFICATION Results (Persent) on KTH-TIPS Training images per class Methods 10 20 30 40 PBR 87.7 ± 2.1 94.3 ± 1.6 97.3 ± 1.2 98.4 ± 1.0 BD 87.2 ± 2.4 94.1 ± 1.8 97.2 ± 1.2 98.3 ± 1.0 JD 87.1 ± 2.5 94.1 ± 1.8 97.1 ± 1.2 98.4 ± 1.0 χ.sup.2 86.5 ± 2.6 94.0 ± 1.9 97.1 ± 1.1 98.5 ± 1.1 L.sub.1 83.3 ± 2.7 92.3 ± 2.0 95.9 ± 1.3 97.6 ± 1.1 L.sub.2 79.9 ± 2.8 89.9 ± 2.1 94.4 ± 1.5 96.5 ± 1.2 L.sub.0.5 74.9 ± 3.0 80.6 ± 2.1 83.6 ± 2.0 84.4 ± 1.7 L.sub.1-BRD 83.4 ± 2.7 92.3 ± 2.0 95.9 ± 1.3 97.6 ± 1.1 HI 83.7 ± 2.9 92.4 ± 1.9 95.8 ± 1.2 97.5 ± 1.2 Hellinger 84.6 ± 2.8 92.8 ± 1.9 96.3 ± 1.2 97.8 ± 1.1

TABLE-US-00010 TABLE 7 TEXTURE CLASSIFICATION Results (Persent) on UMD Training images per class Methods 5 10 15 20 PBR 90.4 ± 2.0 95.6 ± 1.2 97.4 ± 0.9 98.4 ± 0.8 BD 90.1 ± 2.4 95.2 ± 1.2 97.1 ± 1.0 98.3 ± 0.7 JD 90.0 ± 2.3 95.5 ± 1.4 97.3 ± 1.0 98.3 ± 0.8 χ.sup.2 89.8 ± 2.5 95.5 ± 1.4 97.3 ± 1.0 98.3 ± 0.8 L.sub.1 86.7 ± 2.6 93.5 ± 1.5 95.8 ± 1.0 97.1 ± 0.9 L.sub.2 80.3 ± 2.5 89.7 ± 1.5 93.1 ± 1.3 95.0 ± 1.2 L.sub.0.5 84.9 ± 1.7 90.7 ± 1.3 92.8 ± 1.2 94.3 ± 1.1 L.sub.1-BRD 86.7 ± 2.5 93.5 ± 1.5 95.8 ± 1.0 97.1 ± 0.9 HI 86.6 ± 2.6 93.4 ± 1.5 95.8 ± 1.0 97.0 ± 0.9 Hellinger 86.9 ± 2.4 93.3 ± 1.5 95.7 ± 1.1 97.0 ± 1.0

TABLE-US-00011 TABLE 8 TEXTURE CLASSIFICATION Results (Persent) on Kylberg Training images per class Methods 2 3 4 5 PBR 85.0 ± 2.8 90.7 ± 1.9 93.6 ± 1.7 95.8 ± 1.2 BD 83.9 ± 3.6 89.8 ± 2.5 93.1 ± 2.1 95.5 ± 1.5 JD 83.9 ± 2.6 89.8 ± 2.7 93.1 ± 1.9 95.5 ± 1.4 χ.sup.2 83.1 ± 3.3 89.6 ± 2.4 92.5 ± 2.4 95.0 ± 1.6 L.sub.1 78.5 ± 2.8 84.1 ± 2.8 88.2 ± 2.1 90.4 ± 1.8 L.sub.2 73.6 ± 3.1 79.8 ± 3.0 83.9 ± 2.8 87.4 ± 2.3 L.sub.0.5 76.3 ± 4.6 82.9 ± 2.0 85.9 ± 1.9 88.3 ± 1.5 L.sub.1-BRD 78.5 ± 2.8 84.1 ± 2.8 88.1 ± 2.1 90.4 ± 1.8 HI 78.5 ± 2.4 84.4 ± 2.4 88.0 ± 2.1 90.4 ± 1.8 Hellinger 80.5 ± 2.5 86.0 ± 2.0 89.2 ± 1.9 91.2 ± 1.7

Leaf Recognition

[0123] The Swedish leaf data set contains 15 different Swedish tree species, with 75 images per species. These images exhibit high inter-class similarity and high intra-class geometric and photometric variations. We used the same PRICoLBP configuration as for the texture data sets. It should be noted that we did not use the spatial layout prior information of leaves. Experimental results are shown in Table 9. We observe that PBR yields more accurate results than other distance measures.

TABLE-US-00012 TABLE 9 Recognition Results (Percent) on Swedish Leaf Data Set Training images per class Methods 5 10 25 PBR 95.7 ± 1.0 97.7 ± 0.7 99.7 ± 0.2 BD 94.5 ± 1.7 97.2 ± 1.1 99.6 ± 0.3 JD 93.9 ± 1.9 97.2 ± 0.7 99.6 ± 0.3 χ.sup.2 94.0 ± 1.9 97.2 ± 0.9 99.6 ± 0.3 L.sub.1 92.2 ± 1.5 95.8 ± 1.2 98.8 ± 0.6 L.sub.2 85.8 ± 3.5 91.7 ± 1.8 97.2 ± 0.9 L.sub.0.5 88.7 ± 1.5 92.2 ± 1.0 94.8 ± 0.6 L.sub.1-BRD 92.2 ± 1.5 95.8 ± 1.2 98.8 ± 0.6 HI 91.6 ± 1.6 95.6 ± 1.3 98.8 ± 0.6 Hellinger 93.3 ± 1.4 96.7 ± 0.9 99.2 ± 0.5

Material Recognition

[0124] The Flickr Material Database (FMD) is a recently published challenging benchmark data set for material recognition. The images in this database are manually selected from Flickr photos and each image belongs to either one of 10 common material categories, including fabric, foliage, glass, leather, metal, paper, plastic, stone, water, and wood. Each category includes 100 images (50 close-up views and 50 object-level views) which captures the appearance variation of real-world materials. Hence these images have large intra-class variations and different illumination conditions. In effect, they are associated with segmentation masks which describe the location of the object. These masks may be used to extract PRICoLBP only from the object regions. Specifically, the 6-template configuration may be used for PRICoLBP, which yielded a 3,540 dimensional feature vector.

[0125] Table 10 shows the dependence of recognition rates on the number of training images per class of the FMD data set. It was observed that the PBR kernel performs best followed by Bhattacharyya distance and Jeffrey divergence.

TABLE-US-00013 TABLE 10 Experimental Results (Percent) on FMD Data Set Training images per class Methods 5 10 20 30 40 50 PBR 30.3 ± 3.9 39.8 ± 2.1 47.4 ± 1.8 51.5 ± 1.5 55.8 ± 1.7 57.3 ± 2.2 BD 28.6 ± 5.7 38.5 ± 2.2 46.7 ± 1.8 50.6 ± 2.0 54.7 ± 1.6 57.1 ± 2.1 JD 27.3 ± 3.9 38.6 ± 2.2 46.6 ± 2.0 51.0 ± 2.4 55.4 ± 1.9 57.0 ± 2.0 χ.sup.2 27.4 ± 4.2 38.0 ± 1.7 45.9 ± 2.0 50.2 ± 2.4 54.5 ± 1.4 56.4 ± 1.8 L.sub.1 25.3 ± 4.9 35.1 ± 1.7 41.3 ± 2.1 43.6 ± 2.0 47.5 ± 1.6 51.3 ± 1.9 L.sub.2 22.0 ± 3.4 28.2 ± 2.9 34.3 ± 1.2 36.6 ± 1.0 39.9 ± 1.4 42.5 ± 2.2 L.sub.0.5 16.3 ± 2.4 18.2 ± 1.4 21.3 ± 1.3 21.3 ± 1.4 23.9 ± 1.3 23.9 ± 1.8 L.sub.1-BRD 25.1 ± 4.9 34.5 ± 2.3 41.1 ± 2.4 43.6 ± 2.0 47.5 ± 1.6 51.2 ± 1.9 HI 27.1 ± 4.1 35.0 ± 2.0 42.0 ± 1.4 44.3 ± 2.0 47.8 ± 1.7 51.3 ± 1.7 Hellinger 28.5 ± 3.4 35.7 ± 1.8 42.4 ± 1.1 44.5 ± 2.0 49.2 ± 1.3 51.9 ± 1.6

[0126] Note that in Table 11, PBR kernel is the top performer in 5 categories out of all 10 categories, compared to other distance measure kernels.

TABLE-US-00014 TABLE 11 Category-Wise Accuracy (Percent) on FMD Data Set Category PBR BD JD χ.sup.2 HI Fabric 46.6 47.0 46.6 46.2 39.2 Foliage 86.2 85.2 84.2 83.0 78.2 Glass 59.2 59.4 61.0 58.6 42.8 Leather 52.2 52.8 52.6 50.8 54.8 Metal 28.6 28.0 29.2 30.2 24.2 Paper 47.6 45.4 46.6 46.2 46.6 Plastic 52.8 51.8 50.6 49.8 48.4 Stone 69.2 71.2 68.8 70.2 63.0 Water 70.0 69.8 69.8 69.6 61.2 Wood 61.0 60.8 60.4 59.6 54.2

Scene Classification

[0127] The Scene-15 data set contains a total of 4,485 images which is a combination of several earlier data sets [71],[77],[78]. Each image in this data set belongs to one of 15 categories, including bedroom, suburb, industrial, kitchen, living room, coast, forest, highway, inside city, mountain, open country, street, tall building, office, and store. The number of images per category varies from 210 to 410. These images are of different resolutions, hence we resized the images to have the minimum dimension of 256 pixels (while maintaining the aspect ratio).

[0128] We used 2.sub.a template configuration of the PRICoLBP but with two scales (radius of neighbors: 1,2). Hence the dimensionality of the feature vector is 2,360. Table 12 shows the clasification results of the different methods for varying number of training images. We observe that PBR works best with a lower number of training images and yields comparable performance with 100 training images per class.

TABLE-US-00015 TABLE 12 Classification Results (Percent) on Scene-15 Data Set Training images per class Methods 10 20 30 100 PBR 62.4 ± 1.9 69.6 ± 1.4 72.4 ± 1.0 79.7 ± 0.5 BD 61.3 ± 2.0 69.1 ± 1.5 72.0 ± 1.1 79.9 ± 0.4 JD 61.4 ± 1.7 69.1 ± 1.4 71.7 ± 1.0 79.9 ± 0.6 χ.sup.2 61.6 ± 1.3 68.6 ± 1.8 71.9 ± 1.1 79.9 ± 0.7 L.sub.1 58.1 ± 1.2 65.6 ± 0.6 69.1 ± 0.8 77.4 ± 0.5 L.sub.2 50.2 ± 2.7 58.2 ± 1.9 62.4 ± 1.7 73.0 ± 0.6 L.sub.0.5 43.3 ± 1.5 48.4 ± 1.6 50.7 ± 2.0 52.9 ± 5.6 L.sub.1-BRD 58.2 ± 1.1 65.6 ± 0.6 69.1 ± 0.8 77.4 ± 0.5 HI 58.1 ± 1.0 65.2 ± 1.2 69.1 ± 0.8 77.3 ± 0.6 Hellinger 59.4 ± 1.4 66.8 ± 1.0 69.9 ± 1.0 78.6 ± 0.7

Object Recognition

[0129] Caltech-101 data set is an important benchmark data set for object recognition. This contains 9,144 images under 102 categories (101 diverse classes and one background class). The number of images per class varies from 31 to 800. These images exhibit high intra-class variation and they also vary in dimensions. Hence, the images were resized to have the minimum dimension of 256 pixels (while maintaining the aspect ratio). 6 template configurations of the PRICoLBP were used along with two scales (radius of neighbors: 1,2), which results in 7,080 dimension feature.

[0130] Table 13 shows the recognition accuracy of the different methods for varying number of training images. One can observe that the results of the PBR distance kernel is comparable to other distance measure based kernels.

TABLE-US-00016 TABLE 13 Recognition Results (Percent) on Caltech-101Data Set Training images per class Methods 10 20 30 PBR 36.42 44.08 49.16 BD 36.27 43.88 48.73 JD 36.33 43.98 49.15 χ.sup.2 36.22 44.03 49.10 L.sub.1 31.71 39.00 44.18 L.sub.2 24.58 31.27 35.65 L.sub.1-BRD 31.70 39.00 44.18 HI 31.64 38.96 44.07 Hellinger 32.99 40.54 45.88

Species Recognition

[0131] The Leeds Butterfly data set consists of 832 images in total for 10 categories (species) of butterflies. The number of images in each category ranges from 55 to 100. They vary in terms of illumination, pose and dimensions. The images were resized to have the minimum dimension of 256 pixels (while maintaining the aspect ratio). The same setting of PRICoLBP was used as for the texture data sets. Table 14 shows the recognition accuracy of the different methods on the Leeds Butterfly data set for a varying number of training images. It can be observed that PBR kernel achieves comparable performance compared to other distance measure based kernels.

TABLE-US-00017 TABLE 14 Recognition Results (Percent) on Leeds Butterfly Data Set Training images per class Methods 10 20 30 40 PBR 90.8 ± 1.6 94.5 ± 0.8 95.7 ± 0.7 97.5 ± 0.9 BD 90.2 ± 1.6 94.0 ± 1.1 95.0 ± 0.6 97.0 ± 0.8 JD 89.7 ± 2.4 94.4 ± 0.8 95.5 ± 0.6 97.2 ± 1.0 χ.sup.2 89.7 ± 2.4 94.4 ± 1.2 95.6 ± 0.8 97.5 ± 0.8 L.sub.1 88.3 ± 1.3 92.9 ± 1.2 94.9 ± 0.9 96.6 ± 1.4 L.sub.2 82.6 ± 2.3 88.5 ± 1.1 91.6 ± 1.3 93.7 ± 1.9 L.sub.0.5 68.9 ± 2.3 73.4 ± 2.2 74.7 ± 2.2 77.1 ± 1.8 L.sub.1-BRD 88.3 ± 1.2 92.9 ± 1.2 94.9 ± 0.9 96.6 ± 1.4 HI 88.0 ± 2.4 93.0 ± 1.0 95.0 ± 0.9 96.7 ± 1.4 Hellinger 89.1 ± 1.2 93.2 ± 1.2 94.4 ± 1.2 96.8 ± 1.2

[0132] Thus, a number of preferred embodiments have been fully described above with reference to the drawing figures. According to aspects of the present invention, systems and methods are provided that can improve the computational efficiency, speed and accuracy of image recognition systems. Applications of the present invention include medical systems such as medical diagnosis machines, DNA sequencing machines, surgical robots, and other imaging systems. Other applications could include machines for verifying biometric signature, criminal investigation systems, such as finger print identification systems or face recognition systems. The skilled person may recognize other new and useful applications of the above-described inventions.

[0133] Although the invention has been described based upon these preferred embodiments, it would be apparent to those of skill in the art that certain modifications, variations, and alternative constructions could be made to the described embodiments within the spirit and scope of the invention.

[0134] For example, users could be classified by, for example, user profiles, and matching could be limited to users having a specified user profile.

REFERENCES

[0135] The following publicly available publications were referenced above by number [#] and form part of the application. The relevant contents of which are hereby incorporated by reference, which should be readily understood from the context and manner of the reference.

[0136] [1] N. Sebe, M. S. Lew, and D. P. Huijsmans, “Toward Improved Ranking Metrics,” Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 22, no. 10, pp. 1132-1143, 2000.

[0137] [2] W. Dong, L. Huchuan, and Y. Ming-Hsuan, “Least Soft-Threshold Squares Tracking,” in Proc. IEEE Conference on Computer Vision and Pattern Recognition, 23-28 June 2013 2013, pp. 2371-2378.

[0138] [3] G. Healey and R. Kondepudy, “Radiometric CCD Camera Calibration and Noise Estimation,” Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 16, no. 3, pp. 267-276, March 1994.

[0139] [4] J. R. Janesick, Scientific Charge-Coupled Devices. Bellingham, Wash.: SPIE, 2001.

[0140] [5.] C.-H. Lin, J.-S. Tsai, and C.-T. Chiu, “Switching Bilateral Filter With a Texture/Noise Detector for Universal Noise Removal,” Image Processing, IEEE Transactions on, vol. 19, no. 9, pp. 2307-2320, 2010.

[0141] [6] C. Liu, R. Szeliski, S. B. Kang, C. L. Zitnick, and W. T. Freeman, “Automatic Estimation and Removal of Noise from a Single Image,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 30, no. 2, pp. 299-314, 2008.

[0142] [7] N. Young and A. Evans, “Spatio-Temporal Attribute Morphology Filters for Noise Reduction in Image Sequences,” in Proc. International Conference on Image Processing, vol. 1, 2003, pp. I-333-6.

[0143] [8] P. H. Westfall and K. S. S. Henning, Understanding Advanced Statistical Methods. Boca Raton, Fla., USA: CRC Press, 2013.

[0144] [9] A. Tversky and I. Gati, “Similarity, Separability, and the Triangle Inequality,” Psychological review, vol. 89, no. 2, p. 123, 1982.

[0145] [10] W. J. Scheirer, M. J. Wilber, M. Eckmann, and T. E. Boult, “Good Recognition is Non-Metric,” Computing Research Repository, vol. abs/1302.4673, 2013.

[0146] [11] K. Beyer, J. Goldstein, R. Ramakrishnan, and U. Shaft, “When Is “Nearest Neighbor” Meaningful?” in Database Theory ICDT99, ser. Lecture Notes in Computer Science, C. Beeriand P. Buneman, Eds. Springer Berlin Heidelberg, 1999, vol. 1540, pp. 217-235.

[0147] [12] C. Aggarwal, A. Hinneburg, and D. Keim, “On the Surprising Behavior of Distance Metrics in High Dimensional Space,” in Database Theory ICDT 2001, ser. Lecture Notes in Computer Science, J. Bussche and V. Vianu, Eds. Springer Berlin Heidelberg, 2001, vol. 1973, pp. 420-434.

[0148] [13] M. Donahue, D. Geiger, R. Hummel, and T.-L. Liu, “Sparse Representations for Image Decomposition with Occlusions,” in Proc. IEEE Computer Society Conference on Computer Vision and Pattern Recognition, June 1996, pp. 7-12.

[0149] [14] D. W. Jacobs, D. Weinshall, and Y. Gdalyahu, “Classification with Nonmetric Distances: Image Retrieval and Class Representation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 22, no. 6, pp. 583-600, 2000.

[0150] [15] P. Howarth and S. Rger, “Fractional Distance Measures for Content-Based Image Retrieval,” in Advances in Information Retrieval, ser. Lecture Notes in Computer Science, D. Losada and J. Fernndez-Luna, Eds. Springer Berlin Heidelberg, 2005, vol. 3408, pp. 447-456.

[0151] [16] K.-S. Goh, B. Li, and E. Chang, “DynDex: A Dynamic and Non-metric Space Indexer,” in Proc. Tenth ACM International Conference on Multimedia. New York, N.Y., USA: ACM, 2002, pp. 466-475.

[0152] [17] Y. Rubner, J. Puzicha C. Tomasi, and J. M. Buhmann, “Empirical Evaluation of Dissimilarity Measures for Color and Texture,” Computer Vision and Image Understanding, vol. 84, no. 1, pp. 25-43, 2001.

[0153] [18] A. Marzal and E. Vidal, “Computation of Normalized Edit Distance and Applications,” Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 15, no. 9, pp. 926-932, 1993.

[0154] [19] L. Le Cam, “An approximation theorem for the poisson binomial distribution,” Pacific Journal of Mathematics, vol. 10(4), pp. 1181-1197, 1960.

[0155] [20] H. Shen, N. Zamboni, M. Heinonen, and J. Rousu, “Metabolite Identification through Machine Learning—Tackling CASMI Challenge Using FingerID,” Metabolites, vol. 3, no. 2, pp. 484-505, 2013.

[0156] [21] A. C. W. Lai, A. N. N. Ba, and A. M. M6ses, “Predicting Kinase Substrates Using Conservation of Local Motif Density,” Bioinformatics, vol. 28, no. 7, pp. 962-969, 2012.

[0157] [22] A. Niida, S. Imoto, T. Shimamura, and S. Miyano, “Statistical Model-Based Testing to Evaluate the Recurrence of Genomic Aberrations,” Bioinformatics, vol. 28, no. 12, pp. i115-i120, 2012.

[0158] [23] J.-B. Cazier, C. C. Holmes, and J. Broxholme, “GREVE: Genomic Recurrent Event ViEwer to Assist the Identification of Patterns Across Individual Cancer Samples,” Bioinformatics, vol. 28, no. 22, pp. 2981-2982, 2012.

[0159] [24] H. Zhou, M. E. Sehi, J. S. Sinsheimer, and K. Lange, “Association Screening of Common and Rare Genetic Variants by Penalized Regression,” Bioinformatics, vol. 26, no. 19, pp. 2375-2382, 2010.

[0160] [25] A. Wilm, P. P. K. Aw, D. Bertrand, G. H. T. Yeo, S. H. Ong, C. H. Wong, C. C. Khor, R. Petric, M. L. Hibberd, and N. Nagarajan, “LoFreq: a Sequence-Quality Aware, Ultra-Sensitive Variant Caller for Uncovering Cell-Population Heterogeneity from High-Throughput Sequencing Datasets,” Nucleic Acids Research, vol. 40, no. 22, pp. 11 189-11 201, 2012.

[0161] [26] A. S. Macdonald, Encyclopedia of Actuarial Science, J. L. Teugels and B. Sundt, Eds. John Wiley & Sons, Ltd, Chichester, 2004.

[0162] [27] H. U. Gerber, “A Proof of the Schuette-Nesbitt Formula for Dependent Events,” Actuarial Research Clearing House, vol. 1, pp. 9-10, 1979.

[0163] [28] Y. Hwang, J.-S. Kim, and I.-S. Kweon, “Difference-Based Image Noise Modeling Using Skellam Distribution,” Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 34, no. 7, pp. 1329-1341, July 2012.

[0164] [29] J. Darroch, “On the Distribution of the Number of Successes in Independent Trials,” The Annals of Mathematical Statistics, vol. 35, pp. 1317-1321, 1964.

[0165] [30] J.-B. Baillon, R. Cominetti, and J. Vaisman, “A Sharp Uniform Bound for the Distribution of Sums of Bernoulli Trials,” arXiv preprint arXiv:OarX.2350v4, 2013.

[0166] [31] V. N. Gudivada and V. V. Raghavan, “Content Based Image Retrieval Systems,” Computer, vol. 28, no. 9, pp. 18-22, 1995.

[0167] [32] M. Arakeri and G. Ram Mohana Reddy, “An Intelligent Content-Based Image Retrieval System for Clinical Decision Support in Brain Tumor Diagnosis,” International Journal of Multimedia Information Retrieval, vol. 2, no. 3, pp. 175-188, 2013.

[0168] [33] J. Kalpathy-Cramer and W. Hersh, “Automatic Image Modality Based Classification and Annotation to Improve Medical Image Retrieval,” Stud Health Technol Inform, vol. 129, no. Pt 2, pp. 1334-8, 2007.

[0169] [34] B. Marshall, “Discovering Robustness Amongst CBIR Features,” International Journal of Web & Semantic Technology (IJWesT), vol. 3, no. 2, pp. 19-31, April 2012.

[0170] [35] O. Boiman, E. Shechtman, and M. Irani, “In Defense of Nearest-Neighbor Based Image Classification,” in Proc. IEEE Conf. Computer Vision and Pattern Recognition, June 2008, pp. 1-8.

[0171] [36] W. Zhang, J. Sun, and X. Tang, “Cat Head Detection—How to Effectively Exploit Shape and Texture Features,” in Proc. Of European Conf. Computer Vision, 2008, pp. 802-816.

[0172] [37] Z. Weiwei, S. Jian, and T. Xiaoou, “From Tiger to Panda: Animal Head Detection,” Image Processing, IEEE Transactions on, vol. 20, no. 6, pp. 1696-1708, 2011

[0173] [38] T. Kozakaya, S. Ito, S. Kubota, and O. Yamaguchi, “Cat Face Detection with Two Heterogeneous Features,” in Proc. IEEE International Conference on Image Processing, 2009, pp. 1213-1216.

[0174] [39] H. Bo, “A Novel Features Design Method for Cat Head Detection,” in Artificial Intelligence and Computational Intelligence, ser. Lecture Notes in Computer Science. Springer Berlin Heidelberg, 2010, vol. 6319, ch. 47, pp. 397-405.

[0175] [40] G. B. Huang, M. Ramesh, T. Berg, and E. Learned-Miller, “Labeled Faces in the Wild: A Database for Studying Face Recognition in Unconstrained Environments,” University of Massachusetts, Amherst, Tech. Rep. 07-49, October 2007.

[0176] [41] N. Dalai and B. Triggs, “Histograms of Oriented Gradients for Human Detection,” in Proc. IEEE Computer Society Conference on Computer Vision and Pattern Recognition, vol. 1, 2005, pp. 886-893.

[0177] [42] Y. Mitani and Y. Hamamoto, “A Local Mean-Based Nonparametric Classifier,” Pattern Recognition Letters, vol. 27, no. 10, pp. 1151-1159, 2006.

[0178] [43] P. Dollar, C. Wojek, B. Schiele, and P. Perona, “Pedestrian Detection: An Evaluation of the State of the Art,” Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 34, no. 4, pp. 743-761, 2012.

[0179] [44] A. Kembhavi, D. Harwood, and L. S. Davis, “Vehicle Detection Using Partial Least Squares,” Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 33, no. 6, pp. 1250-1265, 2011.

[0180] [45] M. Kaaniche and F. Br'emond, “Recognizing Gestures by Learning Local Motion Signatures of HOG Descriptors,” Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 34, no. 11, pp. 2247-2258, 2012.

[0181] [46] O. Ludwig, D. Delgado, V. Goncalves, and U. Nunes, “Trainable Classifier-Fusion Schemes: An Application to Pedestrian Detection,” in Proc. 12th International IEEE Conference on Intelligent Transportation Systems, 2009, pp. 1-6.

[0182] [47] A. KI″ockner, N. Pinto, Y. Lee, B. Catanzaro, P. Ivanov, and A. Fasih, “PyCUDA and PyOpenCL: A Scripting-Based Approach to GPU Run-Time Code Generation,” Parallel Computing, vol. 38, no. 3, pp. 157-174, 2012.

[0183] [48] O. Chapelle, V. Vapnik, O. Bousquet, and S. Mukherjee, “Choosing Multiple Parameters for Support Vector Machines,” Machine Learning, vol. 46, no. 1-3, pp. 131-159, 2002.

[0184] [49] F. Friedrichs and C. Igel, “Evolutionary Tuning of Multiple SVM Parameters,” Neurocomputing, vol. 64, no. 0, pp. 107-117, 2005.

[0185] [50] S.-W. Lin, Z.-J. Lee, S.-C. Chen, and T.-Y. Tseng, “Parameter Determination of Support Vector Machine and Feature Selection Using Simulated Annealing Approach,” Applied Soft Computing, vol. 8, no. 4, pp. 1505-1512, 2008.

[0186] [51] E. Fix and J. Hodges Jr, “Discriminatory Analysis, Nonparametric Discrimination: Consistency Properties,” USAF School of Aviation Medicine, Randolph Field, Tex., Project 21-49-004, Rept. 4, Contract AF41 (128)-31, Tech. Rep., February 1951.

[0187] [52] X. Wu, V. Kumar, J. Ross Quinlan, J. Ghosh, Q. Yang, H. Motoda, G. McLachlan, A. Ng, B. Liu, P. Yu, Z.-H. Zhou, M. Steinbach, D. Hand, and D. Steinberg, “Top 10 Algorithms in Data Mining,” Knowledge and Information Systems, vol. 14, no. 1, pp. 1-37, 2008.

[0188] [53] K. Fukunaga, Introduction to Statistical Pattern Recognition (2.sup.nd ed.). San Diego, Calif., USA: Academic Press Professional, Inc., 1990.

[0189] [54] A. K. Ghosh, “On Optimum Choice of k in Nearest Neighbor Classification,” Computational Statistics & Data Analysis, vol. 50, no. 11, pp. 3113-3123, 2006.

[0190] [55] G. Toussaint, “Bibliography on Estimation of Misclassification,” Information Theory, IEEE Transactions on, vol. 20, no. 4, pp. 472-479, 1974.

[0191] [56] B. Dasarathy, Nearest Neighbor (NN) Norms: NN Pattern Classification Techniques. Washington: IEEE Computer Society, 1991.

[0192] [57] V. Mnih, “CUDAMat: A CUDA-Based Matrix Class for Python,” Technical Report UTML TR 2009-004, Department of Computer Science, University of Toronto, Tech. Rep., November 2009.

[0193] [58] K. Chang, K. W. Bowyer, S. Sarkar, and B. Victor, “Comparison and Combination of Ear and Face Images in Appearance-Based Biometrics,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 25, pp. 1160-1165, 2003.

[0194] [59] M. Burge and W. Burger, “Ear Biometrics in Computer Vision,” in Proc. 15th International Conference on Pattern Recognition, vol. 2, 2000, pp. 822-826 vol. 2.

[0195] [60] H. Galbally and A. Fierrez, “On the Vulnerability of Fingerprint Verification Systems to Fake Fingerprints Attacks,” in Proc. 40th Annual IEEE International Carnahan Conferences Security Technology, USA, 2006, pp. 130-136.

[0196] [61] A. Iannarelli, Ear Identification. California: Paramount Publishing Company, 1989.

[0197] [62] H. Nejati, L. Zhang, T. Sim, E. Martinez-Marroquin, and G. Dong, “Wonder Ears: Identification of Identical Twins from Ear Images,” in Proc. 21st International Conference on Pattern Recognition, November 2012, pp. 1201-1204.

[0198] [63] A. Kumar and C. Wu, “Automated Human Identification Using Ear Imaging,” Pattern Recognition, vol. 45, no. 3, pp. 956-968, 2012.

[0199] [64] R. Bolle, J. Connell, S. Pankanti, N. Ratha, and A. Senior, “The Relation Between the ROC Curve and the CMC,” in Proc. Fourth IEEE Workshop on Automatic Identification Advanced Technologies, October 2005, pp. 15-20.

[0200] [65] I. Steinwart, D. Hush, and C. Scovel, “Learning from Dependent Observations ,” Journal of Multivariate Analysis, vol. 100, no. 1, pp. 175-194, 2009.

[0201] [66] O. Chapelle, P. Haffner, and V. N. Vapnik, “Support Vector Machines for Histogram-Based Image Classification,” Neural Networks, IEEE Transactions on, vol. 10, no. 5, pp. 1055-1064, 1999.

[0202] [67] P. Brodatz, Textures: A Photographic Album for Artists and Designers. Dover Pubns, 1966.

[0203] [68] E. Hayman, B. Caputo, M. Fritz, and J.-O. Eklundh, “On the Significance of Real-World Conditions for Material Classification,” in Computer Vision-ECCV 2004. Springer, 2004, pp. 253-266.

[0204] [69] Y. Xu, H. Ji, and C. Ferm uller, “Viewpoint Invariant Texture Description using Fractal Analysis,” International Journal of Computer Vision, vol. 83, no. 1, pp. 85-100, 2009.

[0205] [70] G. Kylberg, “The kylberg texture dataset v. 1.0,” Centre for Image Analysis, Swedish University of Agricultural Sciences and Uppsala University, Uppsala, Sweden, External report (Blue series) 35, September 2011. [Online]. Available: http://www.cb.uu.se/gustaf/texture/

[0206] [71] S. Lazebnik, C. Schmid, and J. Ponce, “Beyond Bags of Features: Spatial Pyramid Matching for Recognizing Natural Scene Categories,” in Computer Vision and Pattern Recognition, 2006 IEEE Computer Society Conference on, vol. 2. IEEE, 2006, pp. 2169-2178.

[0207] [72] J. Wang, K. Markert, and M. Everingham, “Learning Models for Object Recognition from Natural Language Descriptions,” in BMVC, vol. 1, 2009, p. 2.

[0208] [73] L. Sharan, R. Rosenholtz, and E. H. Adelson, “Accuracy and Speed of Material Categorization in Real-World Images,” Journal of Vision, vol. 14, no. 10, 2014.

[0209] [74] O. J. O. Soderkvist, “Computer Vision Classification of Leaves from Swedish Trees,” Master's thesis, Linkoping University, SE-581 83 Linkoping, Sweden, September 2001, liTH-ISY-EX-3132.

[0210] [75] L. Fei-Fei, R. Fergus, and P. Perona, “Learning Generative Visual Models from Few Training Examples: An Incremental Bayesian Approach Tested on 101 Object Categories,” Computer Vision and Image Understanding, vol. 106, no. 1, pp. 59-70, 2007.

[0211] [76] X. Qi, R. Xiao, C.G. Li, Y. Qiao, J. Guo, and X. Tang, “Pairwise Rotation Invariant Co-Occurrence Local Binary Pattern,” Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 36, no. 11, pp. 2199-2213, 2014.

[0212] [77] A. Oliva and A. Torralba, “Modeling the Shape of the Scene: A Holistic Representation of the Spatial Envelope,” International journal of computer vision, vol. 42, no. 3, pp. 145-175, 2001.

[0213] [78] L. Fei-Fei and P. Perona, “A Bayesian Hierarchical Model for Learning Natural Scene Categories,” in Computer Vision and Pattern Recognition, 2005. CVPR 2005. IEEE Computer Society Conference on, vol. 2. IEEE, 2005, pp. 524-531.