Image recognition method based on category consistency deep learning

Abstract

The invention provides an image recognition method based on category consistency deep learning. The method comprises the following steps: firstly, marking a training set by using an automatic cooperative positioning method to obtain a category-consistent binary mask; constructing the recognition method by using a feature extraction module, a classifier module and a category-consistent mask learning module; during each iterative training, enabling the feature extraction module to perform feature extraction on the input image; enabling the classifier module to carry out calculation according to the extracted features and give a recognition result; enabling the category-consistent mask learning module to predict a category-consistent binary mask according to the extracted features; calculating a loss value in combination with cross entropy loss and a category-consistent loss function, performing back propagation, and adjusting network parameters of the recognition method, and repeating the steps until the training is finished, and selecting the optimal network parameters as recognition model parameters. By implementing the method, learning of the network on key features can be promoted in a self-supervised learning mode, and high-robustness and high-accuracy image recognition is realized.

Description

technical field [0001] The invention relates to the technical field of image recognition, in particular to an image recognition method based on category consistency deep learning. Background technique [0002] Among image recognition methods, image recognition algorithms based on manual features rely on more manual intervention, and are easily disturbed by factors such as illumination, rotation, and distortion. The convolutional neural network can combine feature extraction and recognition steps. Through end-to-end learning, the convolutional neural network can automatically extract abstract features to achieve accurate and efficient recognition. On the other hand, since the network directly recognizes and trains the entire image during the training process, the irrelevant background in the image prevents the network from focusing on the visual common areas under each category. If the network can be required to locate and segment areas containing visual common features unde...

AI Technical Summary

Problems solved by technology

On the other hand, since the network directly recognizes and trains the entire image during the training process, the irrelevant background in the image prevents the network from focusing on the visual common areas under each category.

Images