Efficient Channel and Bilinear Spatial Attention Module Based on CBAM

The introduction of attention modules can effectively enhance the performance of convolutional neural networks in machine vision tasks. Addressing the issues of partial channel information loss caused by channel dimension reduction and insufficient spatial representation due to overly large convolutional kernels in the Convolutional Block Attention Module (CBAM), this paper proposes a new Efficient Channel and Bilinear Spatial Attention Module (EC-BSAM). Firstly, efficient inter-channel interaction is achieved using convolutional operations without dimension reduction to extract crucial channel attention features. Secondly, based on a strategy of bilinear feature fusion, convolutional operations with two different kernel sizes are employed to extract spatial attention features twice. Feature fusion is achieved through weighted concatenation, enhancing the spatial representation capability of the attention module. Finally, extensive experiments are conducted on the CIFAR-10 image dataset. The experimental results demonstrate that EC-BSAM achieves an improvement of 0.53% and 0.28% in classification accuracy on the backbone networks FastNet_T0 and ResNet8 respectively compared to CBAM. Additionally, EC-BSAM also enhances the classification performance of ResNet18, ShuffleNetV2, and MobileNetV2. The experiments validate that EC-BSAM can effectively enhance the feature extraction capability of convolutional neural networks while maintaining strong compatibility.