LAM: Lightweight Attention Module

Abstract

The attention mechanisms have been widely used in existing methods due to their effectiveness. In the field of computer vision, these mechanisms can be grouped as 1) channel attention mechanisms, which highlight the important channels for images, and 2) spatial attention mechanisms, which focus on the location features for all channels of images. These two groups of mechanisms, which have various strategies for capturing features, actually play complementary roles in image classification. Existing lightweight models based on one group of attention mechanisms have fewer parameters than convolutional networks. However, few works consider their integration and maintain their merits for lightweight neural networks. In this paper, we propose a new Lightweight Attention Module (LAM) for lightweight convolutional neural networks to efficiently integrate these attention mechanisms. Specifically, we use element-wise addition and smaller convolutional kernels in the spatial module, avoiding the vanishing gradient problem. Besides, we replace the multi-layer perceptron (MLP) layer with squeeze-and-excitation layers in the channel module, alleviating the problem of channel dependencies. Finally, we adopt a parallel mechanism to coordinate these two attention modules with low computational complexity. Experimental results on benchmark datasets demonstrate the effectiveness of LAM in terms of image classification tasks, ablation study and robustness analysis.