Research On Fine-grained Image Classification Method Based On Attention Mechanism

With the vigorous development of technology,the repetitive and tedious tasks in people’s daily lives are gradually replaced by many new technologies,which are also applied in many fields,such as access control,face recognition,image search-related information,etc.The accuracy of information recognition has also become a major issue in these fields.Image classification is an important tool to improve the accuracy of information recognition as it can identify image categories based on the features in the images.Fine-grained image classification is an important branch of image classification,and the core of its technique is to use efficient network models or algorithms to distinguish different subcategories under the same category.Therefore,it is of great practical significance and development prospect to explore fine-grained image classification.However,fine-grained image classification has the difficulty of image distinguishing,such as the possibility of recognizing the same subcategory as different categories or different subcategories as the same category,etc.Given the great challenges and important applications of fine-grained image classification,this paper presents two fine-grained image classification methods by investigating three aspects: attention mechanism,multi-scale information fusion,and progressive training:1.A fine-grained classification method based on a mixed attention mechanism.In this paper,we propose a fine-grained classification network based on the Mixed Attention mechanism(MA-Net),which contains a Mixed Attention Module(MAM)based on space and channel,which can effectively focus on the target region in the image,and then obtain fine-grained features with high discriminative power from the key areas.This attention module contains a channel attention sub-module and a spatial attention sub-module,in which the channel attention sub-module compresses the information in the spatial dimension of the image to obtain important information in the channel dimension,and the spatial attention sub-module compresses the information in the channel dimension of the image to obtain important information in the spatial dimension.In addition,to further improve the classification accuracy,the data set is expanded using data augmentation to avoid the overfitting phenomenon of the network.2.Multi-scale-based progressive fine-grained image classification method.In this paper,we propose a Mixed Attention and Multi-Scale and Progressive Network(MAMSP-Net)for image classification at different scales.In addition to the mixed attention mechanism,the proposed method uses different scales of the network to extract specific feature information,and the special progressive training method enables the network to simultaneously learn and fuse feature information from different scales of the image.To motivate the two training steps in the model to work together,the final loss function is obtained by optimizing a combination of four loss functions.In the testing phase of the model,considering that predictions based on different scale information are both valuable and complementary,two types of prediction results are proposed: single prediction results and multiple-output combination prediction results.The proposed multi-output combination prediction results of the MAMSP-Net method achieve accuracy of 88.8%,95.0%,and 93.3% on three international datasets CUB,CAR,and AIR,respectively,which better improves the performance of fine-grained image classification.