Learning Competitive Channel-Wise Attention in Residual Network with Masked Regularization and Signal Boosting

Abstract Image classification is an essential component of expert and intelligent systems. The accuracy and efficiency of image classification algorithms significantly affect the performance of related expert systems. Residual network (ResNet) shows strong superiority in image modeling. However, it has also been proved to be low-efficient. In this study, we proposed a novel channel-wise attention mechanism to alleviate the redundancy of ResNet. We introduce the identity mappings into the scope of channel relationship modeling. In this way, the identity mapping can join the optimized process of self-supplementary modeling. Besides, we present the masked regularization for squeezed signals and enhance the robustness of channel-relation encoding. Finally, we verify the performance of the proposed method. The experiments are carried out on the datasets CIFAR-10, CIFAR-100, SVHN, and ImageNet. The proposed method effectively improves the performance of image classification-related expert systems. Moreover, our approach is hot-swappable, has broad applicability, so it has great practical significance for experts and intelligent systems.