Multi-Attention Mechanism Fusion for Fine-Grained Image Classification

Abstract

In recent years, image classification has developed to the fine-grained level, which has become a new research hotspot. Compared with the traditional image classification task, the fine-grained image classification task has small difficulties due to the influence of image shooting scenes. So focus on mechanism has been widely used in fine-grained image classification problems, but the traditional attention focus on mechanism has the characteristics of first positioning and after processing, the model needs to run step by step and the attention focus on method is single. To further improve the performance of deep convolutional neural networks on the fine-grained image classification task, this paper studies the end-to-end weakly supervised fine-grained image classification model with multiple attention mechanism fusion. In this paper, a fine-grained image deep convolutional network model embedded with four attention focusing mechanisms, they are including: class activation mapping CAM attention focusing method, channel attention CA focusing method, spatial attention SA focusing method and channel spatial confusion attention and CSCA focusing method. On the fine-grained image classification dataset CUB-200-2011, Stanford-dogs, Stanford-cars, the results show that the four attention focusing methods can focus on local features and improve the performance of convolutional network classification performance, among which the channel spatial confusion attention focus method is the most significant improvement in the model classification performance.