Analysis And Research Of Fine-Grained Image Classification Technology Based On Representation Learning

Deep learning method is gradually rising and has been developed on a large scale,machines can understand and use the data provided to it well.Some of the problems in real life are difficult to deal with,which can be realized and solved through deep learning.This paper uses deep learning method to solve the problem of fine-grained classification.Fine-grained image classification is a further classification,that is,region molecular classification.Fine-grained image classification data has the following two characteristics:first,compared with ordinary image classification,it has small inter class spacing and large intra class spacing.Second,the task of fine-grained image classification generally has the problem of less total data.In this paper,we achieve fine-grained image classification task by obtaining better representation,mainly using two ways to obtain distinguishable features and enhance the representation.One is to use attention module to extract and represent representation from the image,the other is to introduce external information to guide representation learning.The first method is to extract and represent the representation from the image using attention module.In this model,the loss function is designed to induce different attention modules to extract discriminative features and confusable features explicitly,and then the original features are added with discriminative features and subtracted from confusable features to enhance distinguishable features and weaken obfuscated features.The representation obtained by feature addition and subtraction is the representation used for classification.The second method is to introduce external information to guide representation learning.We introduce the label description information,the image is difficult to distinguish parts or key areas,the text will have a detailed introduction and related description.Text information and image information are complementary information.Text information described by labels can induce the extraction of image representation.This method obtains image features through CNN and label description features through GCN,transforms classification task into matching task,and calculates the similarity between each image feature and each label description feature.The higher the similarity between image features and label description features under the same label,the better;otherwise,the lower the similarity,the better.In this paper,the first method achieves 88.1%,94.9%and 93.6%accuracy on CUB-200-2011,Stanford cars and FGVC Aircraft datasets respectively,and the specific experimental results are analyzed in 3.5.The second method achieves 90.2%and 92.3%accuracy on CUB-200-2011 and Stanford Dogs datasets respectively.These experimental results show the feasibility of this method.