A method comprises: first, obtaining a to-be-recognized lane line image; then determining, based on the lane line image, a candidate pixel used to recognize a lane line region, to obtain a candidate point set, where the lane line region is a region of a location of a lane line in the lane line image and a surrounding region of the location of the lane line; then selecting a target pixel from the candidate point set, and obtaining at least three location points associated with the target pixel in a neighborhood, where the at least three location points are on one lane line; and finally, performing extension by using the target pixel as a start point and based on the at least three location points associated with the target pixel, to obtain a lane line point set corresponding to the target pixel.

A few-shot urban remote sensing image information extraction method based on meta learning and attention includes building a few-shot urban remote sensing information pre-trained model. During a pre-training stage, pre-training network learning is performed for a few-shot set to fully learn feature information of existing samples and obtain initial feature parameters and a deep convolutional network backbone of the few-shot set; the few-shot urban remote sensing information pre-trained model is a network structure including a convolutional layer, a pooling layer and a fully-connected layer, and includes five sections of convolutional network where each section includes two or three convolutional layers, and an end of each section is connected to one maximum pooling layer to reduce a size of a picture; the number of convolutional kernels inside each section is same, and when closer to the fully-connected layer, the number of convolutional kernels is larger.

A system configured to receive video and/or images from an image capture device over a livestock path, generate feature maps from an image of the video by applying at least a first convolutional neural network, slide a window across the feature maps to obtain a plurality of anchor shapes, determine if each anchor shape contains an object to generate a plurality of regions of interest, each of the plurality of regions of interest being a non-rectangular, polygonal shape, extract feature maps from each region of interest, classify objects in each region of interest, in parallel with classification, predict segmentation masks on at least a subset of the regions of interest in a pixel-to-pixel manner, identify individual animals within the objects based on classifications and the segmentation masks, and count individual animals based on identification, and provide the count to a digital device for display, processing, and/or reporting.

The present disclosure provides a method and apparatus for training a multi-modal data matching degree calculation model, a method and apparatus for calculating a multi-modal data matching degree, an electronic device, a computer readable storage medium and a computer program product, and relates to the field of artificial intelligence technology such as deep learning, image processing and computer vision. The method comprises: acquiring first sample data and second sample data that are different in modalities; constructing a contrastive learning loss function comprising a semantic perplexity parameter, the semantic perplexity parameter being determined based on a semantic feature distance between the first sample data and the second sample data; and training, by using the contrastive learning loss function, an initial multi-modal data matching degree calculation model through a contrastive learning approach, to obtain a target multi-modal data matching degree calculation model.

A model generating apparatus and method are provided. The apparatus receives a plurality of sample images. The apparatus generates a plurality of adversarial samples corresponding to the sample images. The apparatus inputs the sample images and the adversarial samples respectively to a first encoder and a second encoder in a self-supervised neural network to generate a plurality of first feature extractions and a plurality of second feature extractions. The apparatus calculates a similarity of each of the first feature extractions and the second feature extractions to train the self-supervised neural network. The apparatus generates a task model based on the first encoder and a plurality of labeled data.

An image processing method generates a training dataset for training a machine learning model to predict illumination conditions for different positions relative to a scene, the training dataset including training images and reference data. The method includes: obtaining a training image of a training scene acquired by a first camera having an associated first coordinate system; determining local illumination maps associated to a respective position in the training scene in a respective second coordinate system and representing illumination received from different directions around the position; transforming the position of each local illumination map from the second to the first coordinate system; responsive to determining that the transformed position of a local illumination map is visible: transforming the local illumination map from the second to the first coordinate system and including the transformed local illumination map and its transformed position in the reference data associated to the training image.

A novel and useful system and method of improved power performance and lowered memory requirements for an artificial neural network based on packing memory utilizing several structured sparsity mechanisms. The invention applies to neural network (NN) processing engines adapted to implement mechanisms to search for structured sparsity in weights and activations, resulting in a considerably reduced memory usage. The sparsity guided training mechanism synthesizes and generates structured sparsity weights. A compiler mechanism within a software development kit (SDK), manipulates structured weight domain sparsity to generate a sparse set of static weights for the NN. The structured sparsity static weights are loaded into the NN after compilation and utilized by both the structured weight domain sparsity mechanism and the structured activation domain sparsity mechanism. The application of structured sparsity lowers the span of search options and creates a relatively loose coupling between the data and control planes.

Embodiments of the present disclosure provide a method, an electronic device, and a computer program product for training a model. The method may include determining image features, audio features, and text features of a reference object based on reference image information, reference audio information, and reference text information associated with the reference object, respectively. The method may also include constructing a feature tensor from the image features, the audio features, and the text features. In addition, the method may further include decomposing the feature tensor into a first feature vector, a second feature vector, and a third feature vector corresponding to the image features, the audio features, and the text features, respectively, to determine a loss function value of the model. The method may also include updating parameters of the model based on the loss function value.

There is provided a method of re-classifying a clinically significant feature of a medical image as an artifact, comprising: feeding a target medical image captured by a specific medical imaging sensor at a specific setup into a machine learning model, obtaining a target feature map as an outcome of the machine learning model, wherein the target feature map includes target features classified as clinically significant, analyzing the target feature map with respect to sample feature map(s) obtained as an outcome of the machine learning model fed a sample medical image captured by at least one of: the same specific medical imaging sensor and the same specific setup, wherein the sample feature map(s) includes sample features classified as clinically significant, identifying target feature(s) depicted in the target feature map having attributes matching sample feature(s) depicted in the sample feature map(s), and re-classifying the identified target feature(s) as an artifact.

There is provided a method for predicting characteristic information of a target to be recognized. The method comprises: acquiring a plurality of first face images for learning and characteristic information on each first face image; generating a plurality of second face images for learning obtained by synthesizing a mask image with the plurality of first face images for learning by a predetermined algorithm; and training a first neural network by using the plurality of second face images for learning as input data for learning and characteristic information as label data for each second face image corresponding to one of the first face images.