GLCANet: Context Attention for Infrared Small Target Detection

Abstract

Infrared small target detection (IRSTD) refers to extracting small targets from infrared images with noisy interference and blurred background. Due to their small size and low contrast in the image, infrared targets are easily overwhelmed, which requires the network to have a wider receptive field for images and better ability to process local information. How to extract contextual information simply and efficiently remains challenging. In this paper, we propose a global and local context attention network (GLCANet), where the global context extraction module (GCEM) and the local context attention module (LCAM) are devised to address this problem. Specifically, GCEM transforms the feature map from the spatial domain to the frequency domain for feature extraction. Since updating a single value in the frequency domain affects all raw data globally, GCEM enables the network to consider the global context at an early stage and obtain a wider receptive field. LCAM fuses multiple layers of features, where we devise a local context-oriented down-sampling block (LCDB). LCDB transforms the planar dimension of the original feature map into the spatial dimension, which can extract more local contextual information while down-sampling the feature. Experiments on public datasets demonstrate the superiority of our method over representative state-of-the-art IRSTD methods.