Research On Fine-grained Image Classification Based On Deep Residual Network

Fine-grained image classification is a very important and challenging research direction in computer vision research,and it has high research significance in academia and industry.Early image classification,due to its significant image discrimination,often only requires low-level features to complete image classification tasks.However,in daily life,more refined image classification is required.Fine-grained image classification has the characteristics of small differences between classes and large differences within classes.Therefore,the key to fine-grained classification tasks is how to accurately locate the local area where the discriminative features of the object to be classified in the image are located,which is very challenging.The paper analyzes the problem of the loss of fine-grained discriminative feature information in the middle layer when using bilinear convolutional neural network(B-CNN)for fine-grained image classification,and selects the residual network(Res Net)to replace the original B-CNN's two-way convolutional neural network(CNN)to improve the feature extraction ability of the model.And the different layer feature outputs extracted from the last three convolution residual blocks of Res Net are separately bilinear pooling,and the three feature vectors obtained are processed and then spliced for multi-layer feature fusion.The interaction of multi-layer features can enrich the learning of fine-grained features and reduce the loss of feature information in the middle layer.In addition,according to the characteristics of the discriminative feature information of the image is often in the local area,then add attention modules to each residual block of the B-CNN,and the attention mechanism is used to focus on the local discriminative feature information of the image,which can enhance the feature representation ability,and suppress redundant information,so as to more effectively perform fine-grained feature learning.Through comparative experiments,it is verified that the improved bilinear convolutional neural network based on the attention mechanism and multi-layer feature fusion proposed in this paper can effectively improve the accuracy of fine-grained image classification.