Research Of Fine-Grained Image Classification Based On Deep Convolutional Neural Network

In recent years,the level of computer hardware has been continuously improved,which has strongly promoted the development of deep learning.After the breakthrough progress of convolutional neural networks in large-scale image classification tasks,a large number of scholars have devoted themselves to relevant research work.In computer vision tasks,convolutional neural networks are becoming more and more widely used,and the images that need to be processed are gradually becoming fine-grained or even super-granular from large scales.However,those net models that perform well in large-scale image classification tasks are somewhat unsatisfactory in fine-grained image classification tasks.The main reason is that objects in fine-grained images belong to the same large class and have a large degree of similarity.Convolutional neural networks are difficult to extract discriminative features.Therefore,this paper focuses on how to improve the accuracy of fine-grained image classification tasks based on the net of residuals,which has good performance in large-scale image classification.The following are the main progress and achievements.(1)For deep convolutional neural networks,it is difficult to extract a small number of discriminative features from a large number of features in fine-grained images,and a deep attention network method for fine-grained image classification is proposed.The attention mechanism in the field of natural language processing will be fully applied to our fine-grained image classification tasks.By analyzing the characteristic of the convolutional layer's output,the channel dimension and spatial dimension of the feature are output from the convolutional layer respectively.At the two levels,the attention mechanism is applied,and the spatial attention module,the channel attention module and the mix attention module are designed.In addition,ResNext50 is used as an example to embed these three modules of interest in their residual structure units to construct three different deep attention networks,and to publish data sets in Stanford Dogs,CUB200-2011 and Stanford Cars fine-grained images.Experiment on it.The experimental results show that embedding the channel attention module or the spatial attention module in the ResNext50 residual structure unit can improve the fine-grained image classification performance of the network;embedding the mix attention module can greatly improve the fine-grained image classification performance of the network.(2)In order to further improve the classification accuracy of fine-grained images,and based on the deep mix network,which has the best performance in the attention network,and a recurrent deep mix attention network method for fine-grained image classification is proposed.Through the visual analysis of the deep convolution features,the positive correlation between the target position and the depth convolution feature response in the image is clarified.Then,based on this relationship,a method for locating the key regions of the image is proposed.And use the network structure in the RA-CNN to realize the automatic positioning of the network to the key areas of the original image and the cropping and zooming function.Finally,the RA-CNN algorithm is used to realize the net recurrent of the two-way deep mix attention network.Experiments were carried out on the Stanford Dogs,CUB200-2011 and Stanford Cars datasets,which yielded 87.1%,84.9%,and 92.4% test set accuracy,respectively,higher than current FCAN,HIHCA and other algorithms.The experimental results show that the recurrent deep mix attention network is a well fine-grained image classification method,which based on weak supervised information.