The Research Of Fine-Grained Image Categorization Based On Deep Learning

Image classification is one of the research hotspots in the fields of computer vision and machine learning.With the development of computer technology,the existing image classification cannot meet the needs of human beings for different subcategories of the same species.In this context,the task of fine-grained image categorization has attracted extensive attention from both academia and industry.Fine-grained image categorization refers to the classification task of subdividing subcategories under the same species category.At present,although the traditional bag of visual word model has been successfully applied to fine-grained image categorization tasks,it still cannot establish a good mapping relationship between low level features and high level semantics.In addition,because of the similarities between classes and the differences among classes,how to learn the characteristics of the key sub-categories in the target subcategory for fine-grained categorization becomes a difficult point in the current research field.In view of the above-mentioned problems,this paper proposes two kinds of image fine-grained categorization algorithms based on deep learning,the main work is as follows:Firstly,for the problem of image incomplete feature learning due to the similarity between classes and intraclass differences in Fine-grained image categorization,an Fine-grained image categorization method based on multi-layer information fusion is proposed.First,different levels of semantic features can be learned through the VGG-19 model structure.Second,the feature maps of specific layers in the model are fusion with the original input images.The fusion feature maps are fusion as the next layer of similar features.The input of the operation is again fusion to obtain the output of the second-level feature fusion.After similar operations of several layers,the network output of multi-layer information fusion is obtained.Third,feature fusion of the output of the multi-layer information fusion network and the output of the last layer of the VGG-19 model is used as the feature vector of the final input sample.In order to increase the distance between classes and reduce the intraclass distance,a Center Loss function is introduced.And the weighted combination of Center Loss and Softmax is used as the final objective function.Secondly,aiming at the lack of strong feature extraction ability and generalization ability for fine-grained image categorization tasks of existing models,and most models require the problem of labeling key points or key parts of images,we propose a multi-scale spatial attention structure fine-grained image categorization method.Its main ideas are as follows: First,the classical model VGG-19 of the deep convolutional network is adopted as the basic model framework,and different levels of features are extracted.Second,the multi-scale spatial attention mechanism is trained to get the pixel spatial weight feature map;Third,weighted feature maps and original feature maps are spatially weighted using dot multiplication operations to strengthen the spatial feature point information and weaken the interference information.The experimental results on CUB-200-2011,Stanford Dogs and Stanford Cars datasets show that the method has a certain improvement over the classification accuracy of traditional methods.