A lightweight attention mechanism for Infrared Small-Target Detection

Infrared small targets lack obvious texture features, have a small pixel proportion, which make it difficult to be detected, especially for deep learning methods, it is easy to be interfered by the background. In order to detect the infrared small target more accurately, reduce the false detection rate. In this paper, we propose a lightweight attention mechanism module (LAM). This module enhances the selection of channel features by emphasizing the importance of channel, and also maintains the selection and enhancement of spatial features. The deep features are fully preserved and irrelevant background information is filtered out. And the LAM is low parameter number and low computation. At the same time, this paper designs a GELAN structure and integrates it into the network to solve the problem of convergence. Experimental result demonstrates that the model achieves advanced results in NUDT-SIRST infrared small target dataset, obtaining 87.70% of mIoU. Compared with DNA-Net, the model in this paper achieves better results in mIoU, Fa and Pd indexes.