Dynamic convolution with complementary attention for image classification

Convolution is the fundamental unit of operation in convolutional neural networks. However, traditional static convolution is subject to two significant limitations: a weak ability to effectively integrate global and local features and a high computational overhead. This paper introduces an innovative and efficient approach, the Complementary Attention Dynamic Convolution (CADC), to tackle the limitations of traditional static convolution in convolutional neural networks. CADC comprises a three-part architecture: two branches of attention and a fusion module. The dynamic convolution kernels are generated by integrating the results of the two branches of attention, which are weighted and fused. CADC is correlated with the input samples, efficiently integrates global and local features, and generates convolution kernels specific to the spatial location and channel domain. Extensive experiments have been conducted on popular benchmark datasets, and the experimental results demonstrate that the CADC method exhibits excellent performance in terms of the number of model parameters, computational complexity, and classification accuracy.