Research on Improving ResNet18 for Classifying Complex Images Based on Attention Mechanism

The computational resources required for training shallow residual networks are relatively few, but their ability to extract features from images with cluttered backgrounds and unclear feature is limited. This article focused on the relatively shallow residual network ResNet18, and added attention mechanism to improve the network’s performance in learning and classifying complex images. Compared to others who added attention mechanisms to the main structure of the residual module, this article, without changing the main structure design and parameter settings of ResNet18, added the attention mechanism to the residual connection of the residual module to form a new network ResNet18-AM. We designed to add the Channel Attention Module (CAM) to the residual connections that require an increase in the number of feature map channels, in order to enhance the feature expression of important channels; In addition, we designed to add the Spatial Convolution Attention Module (SCAM) on residual connections that do not require an increase in the number of channels, in order to enhance the spatial region features of the feature maps. This article used the pneumonia classification public dataset COVID-19 Radiograph Database for experiments to verify the ability of ResNet18-AM to process complex images. Under the setting of small number of samples per batch and small number of training rounds, it is experimentally proved that the training process converges faster, fluctuates less, and classifies more accurately using the ResNet18 network with the introduction of the attention mechanism.