CNXA: A Novel Attention Mechanism Aided Convolution Network

Abstract

Recently, a new generation of convolutional networks, namely ConvNeXt, is proposed, which has the same number of model parameters as Swin-Transformer and superiority in terms of accuracy. This paper designs a ConvNeXt-base network based on the attention mechanism (CNXA). Specifically, we first embed the channel attention mechanism, the spatial attention mechanism, and the combination of both into ConvNeXt-base to improve the performance of the network to recognize targets in images. Then, a new large kernel-filling attention mechanism is proposed based on the above attention mechanisms. Experiments are finally designed to evaluate the proposed CNXA network. We can see that its TOP-1 accuracy on the ImageNet-100 dataset is about 0.4% higher than that of ConvNeXt-base. Moreover, the experimental part verifies the robustness of the model. Open source code is available in https://github.com/ZJieX/cnxa.