Channel Attention Multi-scale Fusion Network with Attention Weight Propagation Mechanism

As computer vision technology continues to advance swiftly, image classification has emerged as a fundamental task in numerous applications. In complex image classification tasks, Convolutional Neural Networks (CNNs) demonstrate notable benefits over traditional machine learning methods. Although CNN-based attention mechanisms have made substantial progress in classification tasks, they still face limitations in capturing multi-scale features and handling complex scenes. To enhance the association between channel features and multi-scale receptive fields, a novel plug-and-play image classification module-the Channel Attention Multi-scale Fusion Network (CAMF-Net) is proposed. This module optimizes feature fusion strategies through a multi-branch structure, effectively extracting and integrating both local and global features. CAMF-Net combines sparsity, multi-scalability, locality, and globality, incorporating an attention weight propagation mechanism to improve the capture and integration of multi-scale information. Finally, the proposed module’s effectiveness and performance benefits are confirmed through comparative and ablation studies conducted across various benchmark datasets. The experimental findings indicate that CAMF-Net surpasses current methods in both classification accuracy and computational efficiency across a range of benchmark models, highlighting its potential for application in complex visual tasks.