SFMNet:focusing on feature attention for image classification
Wenwei Wu, Xiaonan Luo, Songhua Xu, Ruiai Chen · 2022
In the process of convolution network processing, a convolution of each layer can be regarded as a mapping operation of image features. As we all know, when the human eye maps the prescene capture to the brain neurons, the perception of different parts of a scene is different, especially when the perspective focuses on a part or the characteristics of an object, the visual cortex pays more attention to this part. In the construction of CNN, there is consideration of image perception field, but little consideration of image feature area. Therefore, this paper proposes a convolution module called "select future map" (SFM). The model uses the residual idea of RESNET network and the use of channel attention in SENet network, and develops the attention to its own characteristic attention. The structure of convolution module is divided into three parts. The first part is multi module feature convolution, which is used to extract the features of each module in the image. The second part is feature processing, which is used to pay attention to the proportion of different module features. The third part is feature aggregation, which multiplies the modular convolution and the corresponding specific gravity to obtain the next feature image. Taking this as the basic unit, add modules to form a network, and the final experimental results get an accuracy of 84.12%.