Face detector training method, face detection method, and apparatuses

Abstract

A face detector training method, a face detection method, and apparatuses are provided. In the present invention, during a training phase, a flexible block based local binary pattern feature and a corresponding second classifier are constructed, appropriate second classifiers are searched for to generate multiple first classifiers, and multiple layers of first classifiers that are obtained by using a cascading method form a final face detector; and during a detection phase, face detection is performed on a to-be-detected image by using a first classifier or a face detector that is learned during a training process, so that a face is differentiated from a non-face, and a face detection result is combined and output.

Claims

1. A face detector training method implemented by a computer having a processor, the method comprising: performing, by the processor, a training process by: collecting face and non-face images as a training image sample set; extracting a flexible block based local binary pattern (FBLBP) feature of the face and non-face images to form an FBLBP feature set; and using the FBLBP feature and a GentleBoost algorithm to perform training, to obtain a first classifier, wherein the first classifier comprises several optimal second classifiers, and wherein each optimal second classifier is obtained by training using the GentleBoost algorithm; repeating, by the processor, the training process to obtain multiple layers of first classifiers; and cascading, by the processor, the multiple layers of first classifiers to form a face detector.

2. The method according to claim 1, wherein collecting the face and the non-face images as the training image sample set and extracting the FBLBP feature of the face and the non-face images to form the FBLBP feature set comprises: constructing the FBLBP feature to represent co-occurrence information of a relative change of local grayscale of an image, wherein the FBLBP feature comprises several congruent rectangular blocks, a threshold, and a sign bit, wherein the several congruent rectangular blocks comprise one pivot block and at least one neighbor block, wherein the threshold is θ.sub.0 when the sign bit is −1, wherein the threshold is θ.sub.1 when the sign bit is 1, and wherein θ.sub.0 and θ.sub.1 are real numbers; and calculating a sketch value AvgInt.sub.pivotblock of the pivot block in the FBLBP feature and a sketch value AvgInt.sub.neighborblock.sub.k of each of the neighbor blocks in the FBLBP feature, wherein the sketch value AvgInt.sub.pivotblock of the pivot block is represented by average grayscale of an area in which the pivot block is located, the sketch value AvgInt.sub.neighborblock.sub.k of each of the neighbor blocks is represented by average grayscale of an area in which each of the neighbor blocks is located separately, and k is a sequence number of the k.sup.th neighbor block; calculating a differential result according to the sign bit, and comparing and quantizing the differential result and the threshold for each of the neighbor blocks; converting a binary number that consists of a quantization result of each of the neighbor blocks into a decimal number; and saving the decimal number to obtain a sketch value of the FBLBP feature.

3. The method according to claim 1, wherein in using the FBLBP feature and the GentleBoost algorithm to perform training, to obtain the first classifier, wherein the first classifier comprises the several optimal second classifiers, and each optimal second classifier is obtained by training using the GentleBoost algorithm, a calculation formula (1) of a second classifier is as follows: $f_m\left(x\right)=\{\begin{array}{cc}{a}_0,& x=0\\ a_1,& x=1\\ \mathrm{.Math.}& \\ a_{2^K-1},& x=2^K-1\end{array},$ wherein f.sub.m(x) is the m.sup.th second classifier, x is a sketch value of an FBLBP feature of the second classifier, K is the number of neighbor blocks of the FBLBP feature, and a.sub.j is output of a second classifier, and wherein a.sub.j is calculated according to a formula (2): $a_j=\frac{\underset{i}{.Math.}{w}_i{y}_i\delta \left(x_i=j\right)}{\underset{i}{.Math.}{w}_i\delta \left(x_i=j\right)},$ wherein 0≦j≦2.sup.K−1 and j is an integer, y.sub.i is a class of the i.sup.th training image sample, wherein when the training image sample is a face image, a value of y.sub.i is 1, and when the training image sample is not a face image, the value of y.sub.i is −1, δ( ) is a Kronecker function, wherein when x.sub.i=j is true, output is 1, and when x.sub.i=j is false, output is 0, i is the i.sup.th training image sample, and ω.sub.i is weight of the i.sup.th training image sample.

4. The method according to claim 3, wherein the method further comprises obtaining, by calculation, the weight ω.sub.i of the i.sup.th training image sample according to a formula (3) and a normalization formula (4), wherein the formula (3) is ω.sub.i=ω.sub.i×e.sup.−y.sup.i.sup.f.sup.m.sup.(x.sup.r.sup.), and the formula (4) is ${\omega }_i=\frac{{\omega }_i}{\underset{i}{.Math.}{\omega }_i}.$

5. The method according to claim 3, wherein the method further comprises obtaining, by calculation, a sketch value x of an FBLBP feature of each second classifier according to a formula (5), wherein the formula (5) is as follows: $x=\underset{k=1}{\overset{K}{.Math.}}{2}^{k-1}{\mathrm{FBLBP}}_{\mathrm{sign},k},$ where K is the number of neighbor blocks; if sign=1, FBLBP.sub.sign,k=δ((AvgInt.sub.pivotblock−AvgInt.sub.neighborblock.sub.k)≦θ.sub.1); and if sign=−1, FBLBP.sub.sign,k=δ((AvgInt.sub.pivotblock−AvgInt.sub.neighborblock.sub.k)≦θ.sub.0); wherein δ( ) is a Kronecker function, wherein when input is true, output is 1, and when input is false, output is 0.

6. The method according to claim 1, wherein in using the FBLBP feature and the GentleBoost algorithm to perform training to obtain the first classifier, wherein the first classifier comprises the several optimal second classifiers, a process of calculating the optimal second classifiers comprises the following substeps: initially, the FBLBP feature comprises only one pivot block and one neighbor block, wherein an FBLBP feature set of the FBLBP feature that consists of the two rectangular blocks may be obtained using Brute force and traversal, for each FBLBP feature in the FBLBP feature set, calculating a sketch value of the FBLBP feature and an output value of a corresponding second classifier, substituting the obtained output value of the second classifier into a formula (6) to obtain an error J of the second classifier, and selecting a second classifier having a smallest value of the error J as the optimal second classifier, wherein the formula (6) is as follows: $J=\underset{i=1}{\overset{N}{.Math.}}{w_i\left(y_i-f_m\left(x_i\right)\right)}^2;$  wherein J is an optimized objective function, that is, a weighted classification error, and N is the total number of training image samples; and after the FBLBP feature that comprises only one pivot block and one neighbor block is determined, increasing the number of neighbor blocks for the FBLBP feature by traversing rectangular blocks, in an image, that have a same size as the pivot block, calculating a value x by substituting the sketch value of the FBLBP feature into a formula (5), calculating the error J by substituting the value x into the formula (6), when J further decreases, incorporating a new neighbor block into a current optimal second classifier, when J does not decrease, stopping incorporation, directly outputting the current feature, and meanwhile updating weight and entering a next GentleBoost cycle, wherein a calculation formula (7) for combining an optimal second classifier of each GentleBoost cycle into the first classifier is as follows:
F(x)=F(x)+f.sub.m(x), wherein F(x) is the first classifier, and F(x) is initialized to 0.

7. The method according to claim 6, further comprising: determining a threshold of the first classifier according to a formula (8), wherein the formula (8) is as follows: $\mathrm{th}=\mathrm{argmax}\left(\underset{i_1\in \mathrm{Face}}{.Math.}\delta \left(F\left(x_{i_1}\right)\ge \mathrm{th}\right)+\underset{i_2\in \mathrm{Non}\text{-}\mathrm{face}}{.Math.}\delta \left(F\left(x_{i_2}\right)<\mathrm{th}\right)\right),$ wherein i.sub.1 is the i.sub.1.sup.th training image sample that comprises a face, i.sub.2 is the i.sub.2.sup.th training image sample that comprises a non-face, and th is the threshold of the first classifier.

8. The method according to claim 2, wherein for each of the neighbor blocks, calculating the differential result according to the sign bit, comparing and quantizing the differential result and the threshold, converting the binary number that consists of the quantization result of each of the neighbor blocks into the decimal number, and saving the decimal number to obtain the sketch value of the FBLBP feature comprises, for each neighbor block, when the sign bit is 1 and a difference between the sketch value AvgInt.sub.pivotblock of the pivot block and a sketch value AvgInt.sub.neighborblock.sub.k of the k.sup.th neighbor block is not less than the threshold θ.sub.1, labeling a position of the neighbor block as 1, and otherwise, when the difference between the sketch value AvgInt.sub.pivotblock of the pivot block and the sketch value AvgInt.sub.neighborblock.sub.k of the k.sup.th neighbor block is less than the threshold θ.sub.1, labeling the position of the neighbor block as 0, or for each of the neighbor blocks, when an opposite number of a difference between the sketch value AvgInt.sub.pivotblock of the pivot block and a sketch value AvgInt.sub.neighborblock.sub.k of the k.sup.th neighbor block is not less than the threshold θ.sub.0, labeling a position of the neighbor block as 1, and otherwise, when the opposite number of the difference between the sketch value AvgInt.sub.pivotblock of the pivot block and the sketch value AvgInt.sub.neighborblock.sub.k of the k.sup.th neighbor block is less than the threshold θ.sub.0, labeling the position of the neighbor block as 0.

9. A face detector training apparatus, comprising: a memory comprising instructions; and a processor coupled to the memory, wherein the instructions cause the processor to be configured to: collect face and non-face images as a training image sample set; extract a flexible block based local binary pattern (FBLBP) feature of the face and non-face images to form an FBLBP feature set; perform training, using the FBLBP feature and using a GentleBoost algorithm, to obtain a first classifier, wherein the first classifier comprises several optimal second classifiers, and wherein each optimal second classifier is obtained by training using the GentleBoost algorithm; repeat a training process to obtain multiple layers of first classifiers; and cascade the multiple layers of first classifiers to form a face detector.

10. The apparatus according to claim 9, wherein the instructions further cause the processor to be configured to: construct the FBLBP feature to represent co-occurrence information of a relative change of local grayscale of an image, wherein the FBLBP feature comprises several congruent rectangular blocks, a threshold, and a sign bit, wherein the several congruent rectangular blocks comprise one pivot block and at least one neighbor block, and when the sign bit is −1, the threshold is θ.sub.0, and when the sign bit is 1, the threshold is θ.sub.1, wherein θ.sub.0 and θ.sub.1 are real numbers; calculate a sketch value AvgInt.sub.pivotblock of the pivot block in the FBLBP feature that is-a sketch value AvgInt.sub.neighborblock.sub.k of each of the neighbor block in the FBLBP feature, wherein the sketch value AvgInt.sub.pivotblock of the pivot block is represented by average grayscale of an area in which the pivot block is located, the sketch value AvgInt.sub.neighborblock.sub.k of each of the neighbor blocks is represented by average grayscale of an area in which each of the neighbor blocks is located separately, and k is a sequence number of the k.sup.th neighbor block, for each of the neighbor blocks; calculate a differential result according to the sign bit; compare and quantize the differential result and the threshold; convert a binary number that consists of a quantization result of each of the neighbor blocks into a decimal number; and save the decimal number to obtain a sketch value of the FBLBP feature.

11. The apparatus according to claim 9, wherein the instructions further cause the processor to be configured to use the FBLBP feature and the GentleBoost algorithm to perform training to obtain the first classifier, wherein the first classifier comprises several optimal second classifiers, and each optimal second classifier is obtained by training using the GentleBoost algorithm, wherein a calculation formula (1) of a second classifier is as follows: $f_m\left(x\right)=\{\begin{array}{cc}{a}_0,& x=0\\ a_1,& x=1\\ \mathrm{.Math.}& \\ a_{2^K-1},& x=2^K-1\end{array},$ wherein f.sub.m(x) is the m.sup.th second classifier, x is a sketch value of an FBLBP feature of the second classifier, K is the number of neighbor blocks of the FBLBP feature, and a.sub.j is output of a second classifier, wherein a.sub.j is calculated according to a formula (2): $a_j=\frac{\underset{i}{.Math.}{w}_i{y}_i\delta \left(x_i=j\right)}{\underset{i}{.Math.}{w}_i\delta \left(x_i=j\right)},$ wherein 0≦j≦2.sup.K−1 and j is an integer; y.sub.i is a class of the i.sup.th training image sample, wherein when the training image sample is a face image, a value of y.sub.i is 1, and when the training image sample is not a face image, the value of y.sub.i is −1, δ( ) is a Kronecker function, wherein when x.sub.i=j is true, output is 1, and when x.sub.i=j is false, output is 0, i is the i.sup.th training image sample, and ω.sub.i is weight of the i.sup.th training image sample.

12. The apparatus according to claim 11, wherein the instructions further cause the processor to be configured to obtain, by calculation, the weight ω.sub.i of the i.sup.th training image sample according to a formula (3) and a normalization formula (4), wherein the formula (3) is ω.sub.i=ω.sub.i×e.sup.−y.sup.i.sup.f.sup.m.sup.(x.sup.i.sup.), and the formula (4) is ${\omega }_i=\frac{{\omega }_i}{\underset{i}{.Math.}{\omega }_i}.$

13. The apparatus according to claim 11, wherein the instructions further cause the processor to be configured to obtain, by calculation, a sketch value x of an FBLBP feature of each second classifier in the formula (1) according to a formula (5), wherein the formula (5) is as follows: $x=\underset{k=1}{\overset{K}{.Math.}}{2}^{k-1}{\mathrm{FBLBP}}_{\mathrm{sign},k},$ where K is the number of neighbor blocks; when sign=1, FBLBP.sub.sign,k=δ((AvgInt.sub.pivotblock−AvgInt.sub.neighborblock.sub.k)≦θ.sub.1); and when sign=−1, FBLBP.sub.sign,k=δ((AvgInt.sub.pivotblock−AvgInt.sub.neighborblock.sub.k)≦θ.sub.0); wherein δ( ) is a Kronecker function, wherein when input is true, output is 1, otherwise, when input is false, output is 0.

14. The apparatus according to claim 9, wherein the FBLBP feature comprises only one pivot block and one neighbor block, wherein an FBLBP feature set of the FBLBP feature that consists of the two rectangular blocks may be obtained using Brute force and traversal, for each FBLBP feature in the FBLBP feature set, wherein the first classifier is configured to: calculate a sketch value of the FBLBP feature and an output value of a corresponding second classifier; substitute the obtained output value of the second classifier into a formula (6) to obtain an error J of the second classifier; select a second classifier having a smallest value of the error J as the optimal second classifier, wherein the formula (6) is as follows: $J=\underset{i=1}{\overset{N}{.Math.}}{w_i\left(y_i-f_m\left(x_i\right)\right)}^2,$ wherein J is an optimized objective function, that is, a weighted classification error, and N is the total number of training image samples; increase, after determining that the FBLBP feature comprises only one pivot block and one neighbor block, the number of neighbor blocks for the FBLBP feature by traversing rectangular blocks, which have a same size as the pivot block, in an image; re-calculate a value x by substituting the sketch value of the FBLBP feature into a formula (5); calculate the error J by substituting the value x into the formula (6), when J further decreases, incorporate a new neighbor block into a current optimal second classifier, when J does not decrease, stop incorporation, directly output the current feature, and meanwhile update weight and enter a next GentleBoost cycle; and combine an optimal second classifier that is obtained in each cycle into the first classifier according to a formula (7), wherein the formula (7) is as follows:
F(x)=F(x)+f.sub.m(x), wherein F(x) is the first classifier, and F(x) is initialized to 0.

15. The apparatus according to claim 14, wherein the instructions further cause the processor to be configured to determine a threshold of the first classifier according to a formula (8), wherein the formula (8) is as follows: $\mathrm{th}=\mathrm{argmax}\left(\underset{i_1\in \mathrm{Face}}{.Math.}\delta \left(F\left(x_{i_1}\right)\ge \mathrm{th}\right)+\underset{i_2\in \mathrm{Non}\text{-}\mathrm{face}}{.Math.}\delta \left(F\left(x_{i_2}\right)<\mathrm{th}\right)\right),$ wherein i.sub.1 is the i.sub.i.sup.th training image sample that comprises a face, i.sub.2 is the i.sub.2.sup.th training image sample that comprises a non-face, and th is the threshold of the first classifier.

16. The apparatus according to claim 10, wherein the instructions further cause the processor to be configured to, for each neighbor block, when the sign bit is 1 and a difference between the sketch value AvgInt.sub.pivotblock of the pivot block and a sketch value AvgInt.sub.neighborblock.sub.k of the k.sup.th neighbor block is not less than the threshold θ.sub.1, label a position of the neighbor block as 1, and otherwise, when the difference between the sketch value AvgInt.sub.pivotblock of the pivot block and the sketch value AvgInt.sub.neighborblock.sub.k of the k.sup.th neighbor block is less than the threshold θ.sub.1, label the position of the neighbor block as 0, or for each of the neighbor blocks, when an opposite number of a difference between the sketch value AvgInt.sub.pivotblock of the pivot block and a sketch value AvgInt.sub.neighborblock.sub.k of the k.sup.th neighbor block is not less than the threshold θ.sub.0, label a position of the neighbor block as 1, and otherwise, when the opposite number of the difference between the sketch value AvgInt.sub.pivotblock of the pivot block and the sketch value AvgInt.sub.neighborblock.sub.k of the k.sup.th neighbor block is less than the threshold θ.sub.0, label the position of the neighbor block as 0.

Description

BRIEF DESCRIPTION OF DRAWINGS

(1) FIG. 1 is a schematic diagram of an MBLBP feature in the prior art;

(2) FIG. 2 is a schematic diagram of an MSOF feature in the prior art;

(3) FIG. 3 is a flowchart of Embodiment 1 of a face detector training method according to the present invention;

(4) FIG. 4 is a schematic diagram of an FBLBP feature according to an embodiment of the present invention;

(5) FIG. 5A is a schematic diagram of an FBLBP feature whose pivot block is separated from a neighbor block according to an embodiment of the present invention;

(6) FIG. 5B is a schematic diagram of an FBLBP feature whose pivot block is adjacent to a neighbor block according to an embodiment of the present invention;

(7) FIG. 5C is a schematic diagram of an FBLBP feature whose pivot block intersects with a neighbor block according to an embodiment of the present invention;

(8) FIG. 6 is a schematic diagram of another FBLBP feature according to an embodiment of the present invention;

(9) FIG. 7A is a labeling diagram of each neighbor block in FIG. 6 when a sign bit is 1;

(10) FIG. 7B is a labeling diagram of each neighbor block in FIG. 6 when a sign bit is 1;

(11) FIG. 8 is a flowchart of face detection performed by using a face detector according to the present invention;

(12) FIG. 9 is a schematic structural diagram of a face detector according to an embodiment of the present invention;

(13) FIG. 10A is a schematic diagram of discrete fractions, in a face detection database (FDDB), of detection performance of a face detection method according to an embodiment of the present invention and detection performance of a face detection method in the prior art;

(14) FIG. 10B is a schematic diagram of continued fractions, in an FDDB, of detection performance of a face detection method according to an embodiment of the present invention and detection performance of a face detection method in the prior art;

(15) FIG. 11 is a schematic structural diagram of Embodiment 1 of a face detector training apparatus according to the present invention;

(16) FIG. 12 is a schematic structural diagram of Embodiment 2 of a face detector training apparatus according to the present invention; and

(17) FIG. 13 is a schematic structural diagram of a face detection apparatus according to the present invention.

DESCRIPTION OF EMBODIMENTS

(18) To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following clearly describes the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. The described embodiments are a part rather than all of the embodiments of the present invention. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

(19) FIG. 3 is a flowchart of Embodiment 1 of a face detector training method according to the present invention. In this embodiment, a face detector training process is mainly a process in which training is performed by using an image sample in a training sample set, an appropriate weak classifier set is found and combined into several strong classifiers, and then, by using a cascading method, multiple layers of strong classifiers form a final face detector. A training phase includes the following steps.

(20) Step 1: Collect face and non-face images as a training sample set, and extract an FBLBP feature of the face and non-face images to form an FBLBP feature set.

(21) In this step, an image that includes a face and an image that does not include a face are collected from the Internet or by using a camera as the training sample set, where an image sample that includes a face may be referred to as a positive sample, and an image sample that does not include a face may be referred to as a negative sample. There may be multiple positive samples and negative samples. A specific quantity may be determined according to a specific requirement. After the training sample set is collected, an FBLBP feature is extracted from face and non-face images to form an FBLBP feature set, and an FBLBP weak classifier that can differentiate face and non-face images is constructed for each FBLBP feature in the FBLBP feature set.

(22) Step 2: Use the FBLBP feature and the GentleBoost algorithm to perform training to obtain a first classifier, where the first classifier includes several optimal second classifiers, and each optimal second classifier is obtained by training by using the GentleBoost algorithm.

(23) Generally, in a process of performing training on a feature by using the GentleBoost algorithm, a second classifier is obtained first, where the second classifier is a classifier of relatively low precision, then several optimal classifiers of relatively high precision are selected from multiple second classifiers, and training is continued on the several optimal classifiers to obtain a second classifier of relatively higher precision, that is, a strong classifier. Unless otherwise stated below, a first classifier is a strong classifier and a second classifier is a weak classifier.

(24) In this step, the GentleBoost algorithm is used to perform feature selection among all FBLBP features that are enumerated in the step 1 to obtain several optimal weak classifiers, and then accumulate the weak classifiers to obtain a strong classifier.

(25) Step 3: Repeat a training process from the step 1 to the step 2 to obtain multiple layers of first classifiers, and cascade the multiple layers of first classifiers to form a face detector.

(26) It should be noted that, for the first classifier at each layer, that is, the strong classifier, both a maximum number of weak classifiers and classification precision of the strong classifier may be set. If a requirement for the number of weak classifiers is not met and the precision does not meet a set requirement, the step 1 is performed again to retrain the samples. Otherwise, if the number of weak classifiers at each layer reaches a maximum value or after the precision meets a requirement, training of first classifiers at a next layer is performed (if it is the last layer, training is completed), so as to avoid that the number of weak classifiers is excessively large during training of last several layers and efficiency is affected. In addition, for a learning process of each FBLBP weak classifier in a strong classifier at each layer, a maximum number of neighbor blocks may be set in a process of adding a neighbor block. If the number of neighbor blocks of a single FBLBP weak classifier does not reach a set maximum value, and classification precision increases after an optimal neighbor block is added, the neighbor block is added; otherwise, adding the neighbor block is stopped and a next GentleBoost cycle is started.

(27) In a face detector training method provided by this embodiment of the present invention, a flexible block based local binary pattern feature and a corresponding weak classifier are constructed, appropriate weak classifiers are found to generate multiple strong classifiers, and multiple layers of strong classifiers that are obtained by using a cascading method form a final face detector. During this process, each FBLBP feature includes one pivot block and at least one neighbor block. The pivot block and the neighbor block are equal in size, and positions of each neighbor block and the pivot block are not strictly limited. Therefore, flexibility is high, robustness is improved, and meanwhile a false detection rate is reduced.

(28) Optionally, in the foregoing Embodiment 1 shown in FIG. 3, the step 1 includes the following substeps.

(29) Step 1.1: Construct the FBLBP feature to represent co-occurrence information of a relative change of local grayscale of an image, where the FBLBP feature includes several congruent rectangular blocks, a threshold, and a sign bit, where the several congruent rectangular blocks include one pivot block and at least one neighbor block; and when the sign bit is −1, the threshold is θ.sub.0, and when the sign bit is 1, the threshold is θ.sub.1, where θ.sub.0 and θ.sub.1 are real numbers.

(30) In this substep, the FBLBP feature, that is used to represent the co-occurrence information of the relative change of the local grayscale of the image is constructed. Each FBLBP feature includes several congruent rectangular blocks, a sign bit, and a threshold corresponding to the sigh bit. For example, when the sign bit is −1, the threshold is θ.sub.0; when the sign bit is 1, the threshold is θ.sub.1, where θ.sub.0 and θ.sub.1 are real numbers. The several rectangular blocks include one pivot block and at least one neighbor block. Referring to FIG. 4, FIG. 4 is a schematic diagram of an FBLBP feature according to this embodiment of the present invention.

(31) As shown in FIG. 4, a pivot block is represented by a rectangular block filled with a light gray color and a neighbor block is represented by an unfilled rectangular block. Each FBLBP feature includes one pivot block and at least one neighbor block. Length and width of the pivot block are equal to length and width of the neighbor block. Positions of each neighbor block and the pivot block are not strictly limited. In this step, a face may be cut out from an image that includes a human face and other sceneries as a positive sample, and this positive sample may be scaled to specified dimensions, for example, 24×24, according to a requirement. In addition, an image that does not include a human face may be directly scaled to specified dimensions as a negative sample. After an image sample that meets a requirement is prepared, according to a preset parameter requirement of an FBLBP feature, all FBLBP features meeting this parameter requirement may be enumerated. An FBLBP parameter requirement includes a position of the pivot block, a size of the pivot block, a maximum number of neighbor blocks, a position of a neighbor block relative to the pivot block, a sign bit, a threshold corresponding to the sigh bit, and the like.

(32) Step 1.2: Calculate a sketch value AvgInt.sub.pivotblock of the pivot block in the FBLBP feature and a sketch value AvgInt.sub.neighborblock.sub.k of or each of the neighbor block in the FBLBP feature, where the sketch value AvgInt.sub.pivotblock of the pivot block is represented by average grayscale of an area in which the pivot block is located, the sketch value AvgInt.sub.neighborblock.sub.k of or each of the neighbor block is represented by average grayscale of each area in which one of the neighbor blocks is located, and k is a sequence number of the k.sup.th neighbor block; for each of the neighbor blocks, calculate a differential result according to the sign bit, and compare and quantize the differential result and the threshold; and convert a binary number that consists of quantization results of all the neighbor blocks into a decimal number, and save the decimal number to obtain a sketch value of the FBLBP feature.

(33) An integral image technology is a matrix representation method that can describe global information. In this step, after the FBLBP feature is constructed, a sketch value of each FBLBP feature is calculated by using the integral image technology. An average grayscale value of each rectangular block in the FBLBP feature is calculated first and a sketch value of the corresponding rectangular block is represented by using the average grayscale value. For example, the sketch value AvgInt.sub.pivotblock of the pivot block is represented by the average grayscale of the area in which the pivot block is located, and the sketch value AvgInt.sub.neighborblock of the neighbor block is represented by the average grayscale of the area in which the neighbor block is located. After a sketch value of each rectangular block is calculated, for each neighbor block, a difference between the sketch value AvgInt.sub.pivotblock of the pivot block and the sketch value AvgInt.sub.neighborblock of the neighbor block is calculated, and a differential result is obtained according to the sign bit. Then the differential result and the threshold corresponding to the sign bit are compared, and then quantized according to a comparison result. For an FBLBP feature, after quantization results of all neighbor blocks of this FBLBP feature are obtained, a binary number that consists of all the quantization results is converted into a decimal number, and the decimal number is used as the sketch value of this FBLBP feature.

(34) Optionally, in the foregoing Embodiment 1 shown in FIG. 3, in the step 2 of using the FBLBP feature and the GentleBoost algorithm to perform training to obtain a first classifier, where the first classifier consists of several optimal second classifiers, and each optimal second classifier is obtained by training by using the GentleBoost algorithm, a calculation formula (1) of a second classifier is as follows:

(35) $f_m\left(x\right)=\{\begin{array}{cc}{a}_0,& x=0\\ a_1,& x=1\\ \mathrm{.Math.}& \\ a_{2^K-1},& x=2^K-1\end{array};$
where, f.sub.m(x) is the m.sup.th second classifier, that is, the m.sup.th weak classifier, x is a sketch value of an FBLBP feature of the weak classifier, K is the number of neighbor blocks of this FBLBP feature, and a.sub.j is output of a weak classifier, where a.sub.j may be calculated according to a formula (2):

(36) $a_j=\frac{\underset{i}{.Math.}{w}_i{y}_i\delta \left(x_i=j\right)}{\underset{i}{.Math.}{w}_i\delta \left(x_i=j\right)};$
where, 0≦j≦.sup.2k−1 and j is an integer; y.sub.i is a class of the i.sup.th training sample, where when the training sample is a face image, a value of y.sub.i is 1, and when the training sample is not a face image, the value of y.sub.i is −1; δ( ) is a Kronecker function, where if x.sub.i=j is true, output is 1, and if x.sub.i=j is false, output is 0; i is the i.sup.th training image sample; and ω.sub.i is weight of the i.sup.th training image sample.

(37) Further, in the foregoing embodiment, the weight ω.sub.i of the i.sup.th training image sample is obtained, by calculation, according to a formula (3) and a normalization formula (4).

(38) The formula (3) is ω.sub.i=ω.sub.i×e.sup.−y.sup.i.sup.f.sup.m.sup.(x.sup.i.sup.).

(39) The formula (4) is

(40) ${\omega }_i=\frac{{\omega }_i}{\underset{i}{.Math.}{\omega }_i}.$

(41) Further, in the foregoing embodiment, a sketch value x of an FBLBP feature of each second classifier in the formula (1) is obtained, by calculation, according to a formula (5). The formula (5) is as follows:

(42) $x=\underset{k=1}{\overset{K}{.Math.}}{2}^{k-1}{\mathrm{FBLBP}}_{\mathrm{sign},k},$
where K is the number of neighbor blocks;
if sign=1, FBLBP.sub.sign,k=δ((AvgInt.sub.pivotblock−AvgInt.sub.neighborblock.sub.k)≦θ.sub.1); and
if sign=−1, FBLBP.sub.sign,k=δ((AvgInt.sub.pivotblock−AvgInt.sub.neighborblock.sub.k)≦θ.sub.0);
where δ( ) is a Kronecker function; when input is true, output is 1; otherwise, when input is false, output is 0.

(43) Optionally, in the foregoing Embodiment 1, in the step 2 of using the FBLBP feature and the GentleBoost algorithm to perform training so as to obtain a first classifier, where the first classifier consists of several optimal second classifiers, a process of calculating the optimal second classifiers includes the following substeps.

(44) Step 2.1: Initially, the FBLBP feature includes only one pivot block and one neighbor block, where an FBLBP feature set of the FBLBP feature that consists of the two rectangular blocks may be obtained by using Brute force and traversal; for each FBLBP feature in the FBLBP feature set, calculate a sketch value of the FBLBP feature and an output value of a corresponding second classifier; substitute the obtained output value of the second classifier into a formula (6) to obtain an error J of the second classifier; and select a second classifier having a smallest value of the error J as the optimal second classifier, where the formula (6) is as follows:

(45) $J=\underset{i=1}{\overset{N}{.Math.}}{w_i\left(y_i-f_m\left(x_i\right)\right)}^2,$
where, J is an optimized objective function, that is, a classification error, and N is the total number of training samples.

(46) In this step, an FBLBP feature that includes only one pivot block and one neighbor block is shown in FIG. 5A, FIG. 5B, and FIG. 5C. FIG. 5A is a schematic diagram of an FBLBP feature whose pivot block is separated from a neighbor block according to this embodiment of the present invention. FIG. 5B is a schematic diagram of an FBLBP feature whose pivot block is adjacent to a neighbor block according to this embodiment of the present invention. FIG. 5C is a schematic diagram of an FBLBP feature whose pivot block intersects with a neighbor block according to this embodiment of the present invention. It may be learned from FIG. 5B that, when the number of neighbor blocks and the number of pivot blocks are both 1, and the neighbor block is adjacent to the pivot block, the FBLBP feature in this embodiment of the present invention is an edge feature of a Haar-Like feature.

(47) Step 2.2: After the FBLBP feature that includes only one pivot block and one neighbor block is determined according to the step 2.1, increase the number of neighbor blocks for the FBLBP feature by traversing, in an image, rectangular blocks that have a same size as the pivot block; re-calculate a value x by substituting the sketch value of the FBLBP feature into the formula (5); calculate the error J by substituting the value x into the formula (6); if J further decreases, incorporate a new neighbor block into a current optimal second classifier; if J does not decrease, stop incorporation, directly output the current feature, and meanwhile update weight and enter a next GentleBoost cycle.

(48) A calculation formula (7) for combining an optimal second classifier of each GentleBoost cycle into the first classifier in the step 2 is as follows:
F(x)=F(x)+f.sub.m(x),
where, F(x) is the first classifier, that is, the strong classifier, and F(x) is initialized to 0.

(49) Further, a threshold of the first classifier may be determined according to a formula (8), where the formula (8) is as follows:

(50) $\mathrm{th}=\mathrm{argmax}\left(\underset{i_1\in \mathrm{Face}}{.Math.}\delta \left(F\left(x_{i_1}\right)\ge \mathrm{th}\right)+\underset{i_2\in \mathrm{Non}\text{-}\mathrm{face}}{.Math.}\delta \left(F\left(x_{i_2}\right)<\mathrm{th}\right)\right);$
where i.sub.1 is the i.sub.1.sup.th training image sample that includes a face; i.sub.2 is the i.sub.2.sup.th training image sample that includes a non-face; and th is the threshold of the strong classifier.

(51) Optionally, in the foregoing Embodiment 1, in the step 1.2, for each of the neighbor blocks, calculating a differential result according to the sign bit, and comparing and quantizing the differential result and the threshold; and finally, converting a binary number that consists of quantization results of all the neighbor blocks into a decimal number, and saving the decimal number to obtain a sketch value of the FBLBP feature includes, for each neighbor block, when the sign bit is 1, if a difference between the sketch value AvgInt.sub.pivotblock of the pivot block and a sketch value AvgInt.sub.neighborblock.sub.k of the k.sup.th neighbor block is not less than the threshold θ.sub.1, labeling a position of the neighbor block as 1, and otherwise, if the difference between the sketch value AvgInt.sub.pivotblock of the pivot block and the sketch value AvgInt.sub.neighborblock.sub.k of the k.sup.th neighbor block is less than the threshold θ.sub.1, labeling the position of the neighbor block as 0; or, for each of the neighbor blocks, when the sign bit is −1, if an opposite number of a difference between the sketch value AvgInt.sub.pivotblock of the pivot block and a sketch value AvgInt.sub.neighborblock.sub.k of the k.sup.th neighbor block is not less than the threshold θ.sub.0, labeling a position of the neighbor block as 1, and otherwise, if the opposite number of the difference between the sketch value AvgInt.sub.pivotblock of the pivot block and the sketch value AvgInt.sub.neighborblock.sub.k of the k.sup.th neighbor block is less than the threshold θ.sub.0, labeling the position of the neighbor block as 0.

(52) The following uses FIG. 6 as an example to describe in detail how to calculate a sketch value of an FBLBP feature in this embodiment of the present invention. Refer to FIG. 6.

(53) FIG. 6 is a schematic diagram of another FBLBP feature according to this embodiment of the present invention. As shown in FIG. 6, this FBLBP feature includes 8 neighbor blocks and one pivot block. When a sign bit is −1, a threshold θ.sub.0 is 2; when a sign bit is 1, a threshold θ.sub.1 is −3.

(54) Referring to FIG. 6, average grayscale of an area in which the pivot block is located is 9, that is, a sketch value AvgInt.sub.pivotblock of the pivot block is 9, and average grayscale of areas in which a neighbor block 1 to a neighbor block 8 are located is respectively 7, 8, 12, 11, 19, 20, 6, and 8, that is, sketch values AvgInt.sub.neighborblock.sub.1 to AvgInt.sub.neighborblock.sub.8 of the neighbor block 1 to the neighbor block 8 are respectively 7, 8, 12, 11, 19, 20, 6, and 8.

(55) When the sign bit is 1, differences between the sketch value AvgInt.sub.pivotblock of the pivot block and the sketch value AvgInt.sub.neighborblock of the neighbor block 1 to the neighbor block 8 are 2, 1, −3, −2, −10, −11, 3, and 1 in sequence. The differences between the sketch value AvgInt.sub.pivotblock of the pivot block and the sketch values AvgInt.sub.neighborblock.sub.1, AvgInt.sub.neighborblock.sub.2, AvgInt.sub.neighborblock.sub.3, AvgInt.sub.neighborblock.sub.4, AvgInt.sub.neighborblock.sub.7, AvgInt.sub.neighborblock.sub.8 of the neighbor block 1, the neighbor block 2, the neighbor block 3, the neighbor block 4, the neighbor block 7, and the neighbor block 8 are not less than the threshold θ.sub.1 (θ.sub.1=−3). Therefore, the neighbor block 1, the neighbor block 2, the neighbor block 3, the neighbor block 4, the neighbor block 7, and the neighbor block 8 are labeled as 1. Whereas, the differences between the sketch value AvgInt.sub.pivotblock of the pivot block and the sketch values AvgInt.sub.neighborblock.sub.5 and AvgInt.sub.neighborblock.sub.6 of the neighbor block 5 and the neighbor block 6 are less than θ.sub.1 (θ.sub.1=−3). Therefore, the neighbor block 5 and the neighbor block 6 are labeled as 0. Referring to FIG. 7A, FIG. 7A is a labeling diagram of each neighbor block in FIG. 6 when the sign bit is 1. According to a sequence of the neighbor block 1 to the neighbor block 8, an obtained binary sequence is 11110011.

(56) When the sign bit is 0, differences between the sketch value AvgInt.sub.pivotblock of the pivot block and the sketch values AvgInt.sub.neighborblock.sub.1 to AvgInt.sub.neighborblock.sub.8 of the neighbor block 1 to the neighbor block 8 are 2, 1, −3, −2, −10, −11, 3, and 1 in sequence, that is, opposite numbers of the differences between the sketch value AvgInt.sub.pivotblock of the pivot block and the sketch values AvgInt.sub.neighborblock.sub.1 to AvgInt.sub.neighborblock.sub.8 of or the neighbor block 1 to the neighbor block 8 are −2, −1, 3, 2, 10, 11, −3, and −1 in sequence. The opposite numbers of the differences between the sketch value AvgInt.sub.pivotblock of the pivot block and the sketch values AvgInt.sub.neighborblock.sub.3, AvgInt.sub.neighborblock.sub.4, AvgInt.sub.neighborblock.sub.5, AvgInt.sub.neighborblock.sub.6 of the neighbor block 3, the neighbor block 4, the neighbor block 5, and the neighbor block 6 are not less than the threshold θ.sub.0 (θ.sub.0=2). Therefore, the neighbor block 3, the neighbor block 4, the neighbor block 5, and the neighbor block 6 are labeled as 1. Whereas, the opposite numbers of the differences between the sketch value AvgInt.sub.pivotblock of the pivot block and the sketch values AvgInt.sub.neighborblock.sub.1, AvgInt.sub.neighborblock.sub.2, AvgInt.sub.neighborblock.sub.7, AvgInt.sub.neighborblock.sub.8 of the neighbor block 1, the neighbor block 2, the neighbor block 7, and the neighbor block 8 are less than θ.sub.0 (θ.sub.0=2). Therefore, the neighbor block 1, the neighbor block 2, the neighbor block 7, and the neighbor block 8 are labeled as 0. Referring to FIG. 7B, FIG. 7B is a labeling diagram of each neighbor block in FIG. 6 when the sign bit is −1. According to the sequence of the neighbor block 1 to the neighbor block 8, an obtained binary sequence is 00111100.

(57) In addition, based on the foregoing face detector training method, an embodiment of the present invention further provides a method for performing face detection by using a face detector that is obtained by using the face detector training method. In this method, face detection is performed on a to-be-detected image by using a first classifier or a face detector that is learned from a training process, so that a face is differentiated from a non-face. Referring to FIG. 8, FIG. 8 is a flowchart of face detection performed by using a face detector according to the present invention, and the flowchart includes the following steps.

(58) Step 4: Traverse a to-be-detected image to obtain a to-be-detected subimage set.

(59) In this step, a preset ratio may be set according to a different scenario. Preset ratios under different scenarios may be different. For example, it is assumed that the to-be-detected image is a 100×200 image; then, 24×24 pixels may be used as a reference to continuously enlarge a detection window (100×100 at maximum, which is equal to a smaller value between length and width of the to-be-detected image) according to a ratio of 1:1.1, an obtained window is used to traverse the to-be-detected image at a step of 2, and obtained sub-windows (24×24, 26×26, . . . , 100×100) of the to-be-detected image are all scaled to a square whose length and width are both 24 pixels, so as to obtain the to-be-detected subimage set. In another possible implementation manner, a preset ratio may also be another value, for example, 1:1.2, which is not limited in the present invention.

(60) Step 5: Bring each to-be-detected subimage in the to-be-detected subimage set into a face detector, and calculate, layer by layer, output of a first classifier at each layer in the face detector.

(61) Step 6: For a to-be-detected subimage, consider that the to-be-detected subimage is a non-face if output of a first classifier at any layer of the face detector is less than a threshold that is of the first classifier and is obtained by training, where only a to-be-detected subimage that passes determining of classifiers at all layers is considered as a face.

(62) In this step, a next layer is entered only after a to-be-detected subimage passes a strong classifier at a previous layer. As a result, a large number of non-target subimages to be detected, for example, a to-be-detected subimage that does not include a face, may be quickly excluded at the first several layers, thereby saving time to perform detection on a target to-be-detected subimage. Referring FIG. 9, FIG. 9 is a schematic structural diagram of a face detector according to an embodiment of the present invention. Referring to FIG. 9, each layer of the face detector corresponds to a strong classifier that is trained by using the GentleBoost algorithm. During detection, it is determined that a to-be-detected subimage is a face image only when the to-be-detected subimage passes determining of all layers, that is, output of the to-be-detected subimage of a strong classifier at each layer is greater than a threshold of this strong classifier. Otherwise, if output of a strong classifier at any layer is less than a threshold of the strong classifier at this layer, a to-be-detected subimage is considered as a non-face image and is immediately abandoned.

(63) Step 7: Combine all detection results in the step 6 to output a position of a face in the to-be-detected image.

(64) A non-maximum suppression method may be used to implement combination of all detection results in the step 6.

(65) According to the face detection method provided by this embodiment of the present invention, face detection is performed on a to-be-detected image by using a strong classifier or a face detector that is learned in a training process, so that a face is differentiated from a non-face, and face detection results are combined and output. During this process, each FBLBP feature includes one pivot block and at least one neighbor block. The pivot block and the neighbor block are equal in size, and positions of each neighbor block and the pivot block are not strictly limited. Therefore, flexibility is high, robustness is improved, and meanwhile a false detection rate is reduced.

(66) Optionally, in the foregoing embodiment shown in FIG. 8, the step 4 includes the following substeps.

(67) Step 4.1: Traverse the to-be-detected image in a detection window with a preset length-width ratio.

(68) Step 4.2: Enlarge the detection window by the preset length-width ratio according to a preset step to traverse the to-be-detected image, and repeat this operation to traverse the to-be-detected image in a different detection window until length of the detection window is greater than or equal to length of the to-be-detected image or until width of the detection window is greater than or equal to width of the to-be-detected image.

(69) Step 4.3: Perform normalization processing on subimages that are obtained by traversing the to-be-detected image in the step 4.1 and the step 4.2, so that a length-width ratio of each subimage conforms to the preset length-width ratio, and use a set that consists of all normalized subimages as the to-be-detected subimage set.

(70) To clearly compare beneficial effects of a face detection method of the present invention with those of a face detection method in the prior art, the following uses Table 1 for a detailed comparison. Table 1 is an effect comparison table of a face detection method provided by the present invention and a face detection method in the prior art.

(71) TABLE-US-00001 TABLE 1 Face Detection Database (Number of Correct Detections/Number of False Detections) Test Dimension (Number Competitor in International of Images) China Competitor FBLBP Age - Baby (113) 104/12 111/5 111/1 Age - Children (105) 104/11 103/3 108/2 Age - Elder (107) 102/13 105/3 108/3 Backlight (145) 139/4  143/2 137/0 Complex Angle (116) 112/2  116/1 116/0 Exp - Frown (92) 92/1  92/0  91/0 Exp - Full smile1 (102) 102/1  102/0 102/0 Exp - Half smile (113) 113/0  110/0 113/0 Exp - Other (63) 63/0  62/1  63/0 Maxnum (4) 69/0  14/0  81/1 No face (32)  0/1  0/1  0/0 Normal (133) 151/1  151/0 151/1 Occa - Glasses (117) 112/5  121/0 121/0 Occa - Hair (55) 50/1  52/0  57/1 Occa - Hand (43) 25/6  23/3  28/1 Occa - Hat (103)  93/20  99/7 101/2 Occa - Sunglasses (99)  75/12  95/12 105/4 Pitch (56) 56/0  56/0  56/0 Race - Black (166) 159/8  176/3 171/2 Race - Brown (70) 70/3  70/0  70/0 Race - White (144) 144/1  143/1 144/1 Rotation - ( ) (111) 110/1  110/1 111/0 Yaw - 22.5 (130) 122/2  128/0 130/1 Total (2219) 2167/105 2182/43 2275/20

(72) It may be learned from Table 1 that, a true positive rate of an FBLBP feature-based face detection method in this embodiment of the present invention is higher than a true positive rate of detection by other international competitors and competitors in China, and a false detection rate is less than a false detection rate of detection by other international competitors and competitors in China.

(73) In addition, to further compare beneficial effects of the face detection method of the present invention and those of the face detection method in the prior art clearly, the following describes a test performed by using an international open FDDB. For a specific result, refer to FIG. 10A and FIG. 10B.

(74) FIG. 10A is a schematic diagram of discrete fractions, in an FDDB, of detection performance of a face detection method according to an embodiment of the present invention and detection performance of a face detection method in the prior art. FIG. 10B is a schematic diagram of continued fractions, in an FDDB, of detection performance of a face detection method according to an embodiment of the present invention and detection performance of a face detection method in the prior art. As shown in FIG. 10A and FIG. 10B, a horizontal coordinate axis is the number of images in a data set that are incorrectly detected by the face detector as faces (False Negative (FN)); a vertical coordinate axis is a true positive rate (TPR); a curve {circle around (1)} represents fractions, in an FDDB, of a detection result of the face detection method (indicated as ours in the figure) provided by this embodiment of the present invention; a curve {circle around (2)} represents fractions, in the FDDB, of a detection result of a jain model; a curve {circle around (3)} represents fractions, in the FDDB, of a detection result of a LiIntel model; a curve {circle around (4)} represents fractions, in the FDDB, of a detection result of a subburaman model; a curve {circle around (5)} represents fractions, in the FDDB, of a detection result of a Viola Jones model; and a curve {circle around (6)} represents fractions, in the FDDB, of a detection result of a Mikolajczyk model. In FIG. 10A, fractions are discrete fractions; and in FIG. 10B, fractions are continued fractions. It may be learned from FIG. 10A and FIG. 10B that, under a certain condition, for example, when the number of positive samples that are predicted as negative samples by the face detector exceeds 10, and a same number of errors is received, the face detection method provided by this embodiment of the present invention has a highest true positive rate.

(75) FIG. 11 is a schematic structural diagram of Embodiment 1 of a face detector training apparatus according to the present invention. The face detector training apparatus provided by this embodiment of the present invention is configured to implement a face detector training method provided by Embodiment 1 of the present invention. For an implementation principle and an effect, refer to the foregoing Embodiment 1 shown in FIG. 3, which is not described herein again. The face detector training apparatus 100 provided by this embodiment of the present invention includes an FBLBP feature set module 11 configured to collect face and non-face images as a training sample set, and extract an FBLBP feature of the face and non-face images to form an FBLBP feature set; a first classifier module 12 configured to perform training, by using the FBLBP feature that is obtained by collection by the FBLBP feature set module 11 and the GentleBoost algorithm, to obtain a first classifier, where the first classifier includes several optimal second classifiers, and each optimal second classifier is obtained by training by using the GentleBoost algorithm; and a face detector module 13 configured to repeat a training process of the FBLBP feature set module 11 and the first classifier module 12 to obtain multiple layers of first classifiers, and cascade the multiple layers of first classifiers to form a face detector.

(76) With a face detector training apparatus provided by this embodiment of the present invention, a flexible block based local binary pattern feature and a corresponding weak classifier are constructed, appropriate weak classifiers are found to generate multiple strong classifiers, and multiple layers of strong classifiers that are obtained by using a cascading method form a final face detector. During this process, each FBLBP feature includes one pivot block and at least one neighbor block. The pivot block and the neighbor block are equal in size, and positions of each neighbor block and the pivot block are not strictly limited. Therefore, flexibility is high, robustness is improved, and meanwhile a false detection rate is reduced.

(77) FIG. 12 is a schematic structural diagram of Embodiment 2 of a face detector training apparatus according to the present invention. As shown in FIG. 12, a face detector training apparatus 200 of this embodiment is based on an apparatus structure shown in FIG. 11, and the FBLBP feature set module 11 includes a constructing unit 111 configured to construct the FBLBP feature to represent co-occurrence information of a relative change of local grayscale of an image, where the FBLBP feature includes several congruent rectangular blocks, a threshold, and a sign bit, where the several congruent rectangular blocks include one pivot block and at least one neighbor block; and when the sign bit is 1, the threshold is θ.sub.0, and when the sign bit is 1, the threshold is θ.sub.1, where θ.sub.0 and θ.sub.1 are real numbers; and a calculating unit 112 configured to calculate a sketch value AvgInt.sub.pivotblock of the pivot block in the FBLBP feature that is constructed by the constructing unit 111 and a sketch value AvgInt.sub.neighborblock.sub.k of each of the neighbor block in the FBLBP feature that is constructed by the constructing unit 111, where the sketch value AvgInt.sub.pivotblock of the pivot block is represented by average grayscale of an area in which the pivot block is located, the sketch value AvgInt.sub.neighborblock.sub.k of each of the neighbor block is represented by average grayscale of an area in which each of the neighbor block is located respectively, and k is a sequence number of the k.sup.th neighbor block; for each of the neighbor block, calculate a differential result according to the sign bit, and compare and quantize the differential result and the threshold; and convert a binary number that consists of quantization results of all the neighbor blocks into a decimal number, and save the decimal number to obtain a sketch value of the FBLBP feature.

(78) Optionally, in an embodiment of the present invention, the first classifier module 12 is configured to use the FBLBP feature and the GentleBoost algorithm to perform training to obtain the first classifier, where the first classifier consists of several optimal second classifiers, and each optimal second classifier is obtained by training by using the GentleBoost algorithm, where a calculation formula (1) of a second classifier is as follows:

(79) $f_m\left(x\right)=\{\begin{array}{cc}{a}_0,& x=0\\ a_1,& x=1\\ \mathrm{.Math.}& \\ a_{2^K-1},& x=2^K-1\end{array},$
where f.sub.m(x) is the m.sup.th second classifier, x is a sketch value of an FBLBP feature of the second classifier, K is the number of neighbor blocks of the FBLBP feature, and a.sub.j is output of a second classifier, where a.sub.j is calculated according to a formula (2):

(80) 0$a_j=\frac{\underset{i}{.Math.}{w}_i{y}_i\delta \left(x_i=j\right)}{\underset{i}{.Math.}{w}_i\delta \left(x_i=j\right)},$
where 0≦j≦2.sup.K−1 and j is an integer; y.sub.i is a class of the i.sup.th training sample, where when the training sample is a face image, a value of y.sub.i is 1, and when the training sample is not a face image, the value of y.sub.i is −1; δ( ) is a Kronecker function, where if x.sub.i=j is true, output is 1, and if x.sub.i=j is false, output is 0; i is the i.sup.th training image sample; and ω.sub.i is weight of the i.sup.th training image sample.

(81) Optionally, in an embodiment of the present invention, the first classifier module 12 is configured to obtain, by calculation, the weight ω.sub.i of the ith training image sample according to a formula (3) and a normalization formula (4), where the formula (3) is ω.sub.i=ω.sub.i×e.sup.−y.sup.i.sup.f.sup.m.sup.(x.sup.i.sup.); and the formula (4) is

(82) ${\omega }_i=\frac{{\omega }_i}{\underset{i}{.Math.}{\omega }_i}.$

(83) Optionally, in an embodiment of the present invention, the first classifier module 12 is configured to obtain, by calculation, a sketch value x of an FBLBP feature of each second classifier in the formula (1) according to a formula (5), where the formula (5) is as follows:

(84) $x=\underset{k=1}{\overset{K}{.Math.}}{2}^{k-1}{\mathrm{FBLBP}}_{\mathrm{sign},k},$
where K is the number of neighbor blocks;
if sign=1, FBLBP.sub.sign,k=δ((AvgInt.sub.pivotblock−AvgInt.sub.neighborblock.sub.k)≦θ.sub.1); and
if sign=−1, FBLBP.sub.sign,k=δ((AvgInt.sub.pivotblock−AvgInt.sub.neighborblock.sub.k)≦θ.sub.0);
where δ( ) is a Kronecker function; when input is true, output is 1; and otherwise, when input is false, output is 0.

(85) Referring to FIG. 12, optionally, in an embodiment of the present invention, the first classifier module 12 includes an initial calculation module 121 configured to initially, the FBLBP feature includes only one pivot block and one neighbor block, where an FBLBP feature set of an FBLBP feature that consists of the two rectangular blocks may be obtained by using Brute force and traversal; for each FBLBP feature in the FBLBP feature set, calculate a sketch value of the FBLBP feature and an output value of a corresponding second classifier; substitute the obtained output value of the second classifier into a formula (6) to obtain an error J of the second classifier; and select a second classifier having a smallest value of the error J as the optimal second classifier; where the formula (6) is as follows:

(86) $J=\underset{i=1}{\overset{N}{.Math.}}{w_i\left(y_i-f_m\left(x_i\right)\right)}^2,$
where J is an optimized objective function, that is, a weighted classification error, and N is the total number of training samples; a cyclic calculation module 122 configured to, after the initial calculation unit 121 determines the FBLBP feature that includes only one pivot block and one neighbor block, increase the number of neighbor blocks for the FBLBP feature by traversing, in an image, rectangular blocks that have a same size as the pivot block; re-calculate a value x by substituting the sketch value of the FBLBP feature into the formula (5); calculate the error J by substituting the value x into the formula (6); if J further decreases, incorporate a new neighbor block into a current optimal second classifier; and if J does not decrease, stop incorporation, directly output the current feature, and meanwhile update weight and enter a next GentleBoost cycle; and a combining unit 123 configured to combine an optimal second classifier that is obtained in each cycle by the cyclic calculation unit 122 into the first classifier according to a formula (7), where the formula (7) is as follows:
F(x)=F(x)+f.sub.m(x),
where F(x) is the first classifier, and F(x) is initialized to 0.

(87) Optionally, referring to FIG. 12 again, in an embodiment of the present invention, the first classifier module 12 further includes a threshold determining unit 124 configured to determine a threshold of the first classifier according to a formula (8), where the formula (8) is as follows:

(88) $\mathrm{th}=\mathrm{argmax}\left(\underset{i_1\in \mathrm{Face}}{.Math.}\delta \left(F\left(x_{i_1}\right)\ge \mathrm{th}\right)+\underset{i_2\in \mathrm{Non}\text{-}\mathrm{face}}{.Math.}\delta \left(F\left(x_{i_2}\right)<\mathrm{th}\right)\right);$
where i.sub.1 is an i.sub.1.sup.th training image sample that includes a face; i.sub.2 is an i.sub.2.sup.th training image sample that includes a non-face; and th is the threshold of the first classifier.

(89) Optionally, in an embodiment of the present invention, the calculating unit 112 is configured to, for each neighbor block, when the sign bit is 1, if a difference between the sketch value AvgInt.sub.pivotblock of the pivot block and a sketch value AvgInt.sub.neighborblock.sub.k of the kth neighbor block is not less than the threshold θ.sub.1, label a position of the neighbor block as 1, and otherwise, if the difference between the sketch value AvgInt.sub.pivotblock of the pivot block and the sketch value AvgInt.sub.neighborblock.sub.k of the k.sup.th neighbor block is less than the threshold θ.sub.1, label the position of the neighbor block as 0; or, for each of the neighbor block, when the sign bit is −1, if an opposite number of a difference between the sketch value AvgInt.sub.pivotblock of the pivot block and a sketch value AvgInt.sub.neighborblock.sub.k of the k.sup.th neighbor block is not less than the threshold θ.sub.0, label a position of the neighbor block as 1, and otherwise, if the opposite number of the difference between the sketch value AvgInt.sub.pivotblock of the pivot block and the sketch value AvgInt.sub.neighborblock.sub.k of the k.sup.th neighbor block is less than the threshold θ.sub.0, label the position of the neighbor block as 0.

(90) FIG. 13 is a schematic structural diagram of a face detection apparatus according to the present invention. The face detection apparatus provided by this embodiment of the present invention is used to implement the face detection method provided by the embodiment shown in FIG. 8 of the present invention. For an implementation principle and an effect, refer to the foregoing FIG. 8, which is not described herein again. The face detection apparatus 300 provided by this embodiment of the present invention includes a traversing module 31 configured to traverse a to-be-detected image so as to obtain a to-be-detected subimage set; a calculating module 32 configured to bring each to-be-detected subimage in the to-be-detected subimage set that is obtained by traversing by the traversing module 31 into the face detector, and calculate, layer by layer, output of a first classifier at each layer in the face detector; a determining module 33 configured to, for each to-be-detected subimage, determine the output that is of the first classifier at each layer and is calculated by the calculating module 32, and consider that the to-be-detected subimage is a non-face if output of a first classifier at any layer of the face detector is less than a threshold that is of the first classifier and is obtained by training, where only a to-be-detected subimage that passes determining of classifiers at all layers is considered as a face; and a combining module 34 configured to combine all detection results that are obtained by the determining module 33, and output a position of a face in the to-be-detected image.

(91) With the face detection apparatus provided by this embodiment of the present invention, face detection is performed on a to-be-detected image by using a strong classifier or a face detector that is learned in a training process, so that a face is differentiated from a non-face, and face detection results are combined and output. During this process, each FBLBP feature includes one pivot block and at least one neighbor block. The pivot block and the neighbor block are equal in size, and positions of each neighbor block and the pivot block are not strictly limited. Therefore, flexibility is high, robustness is improved, and meanwhile a false detection rate is reduced.

(92) Optionally, in the foregoing embodiment, the traversing module 31 is configured to traverse the to-be-detected image in a detection window with a preset length-width ratio; enlarge the detection window by the preset length-width ratio according to a preset step to traverse the to-be-detected image, and repeat this operation to traverse the to-be-detected image in a different detection window until length of the detection window is greater than or equal to length of the to-be-detected image or until width of the detection window is greater than or equal to width of the to-be-detected image; and perform normalization processing on subimages that are obtained by traversing the to-be-detected image as described above, so that a length-width ratio of each subimage conforms to the preset length-width ratio, and use a set that consists of all normalized subimages as the to-be-detected subimage set.

(93) A person of ordinary skill in the art may understand that all or a part of the steps of the method embodiments may be implemented by a program instructing relevant hardware. The program may be stored in a computer readable storage medium. When the program runs, the steps of the method embodiments are performed. The foregoing storage medium includes any medium that can store program code, such as a read-only memory (ROM), a random-access memory (RAM), a magnetic disk, or an optical disc.

(94) Finally, it should be noted that the foregoing embodiments are merely intended for describing the technical solutions of the present invention, but not for limiting the present invention. Although the present invention is described in detail with reference to the foregoing embodiments, persons of ordinary skill in the art should understand that they may still make modifications to the technical solutions described in the foregoing embodiments or make equivalent replacements to some or all technical features thereof, without departing from the scope of the technical solutions of the embodiments of the present invention.