ISPANet: A Pyramid Self-Attention Network for Single-Frame High-Resolution Infrared Small Target Detection With a Large-Scale Dataset SHR-IRST

Abstract

In recent years, the imaging resolution of infrared detection equipment has gradually increased, and higher resolution infrared detection equipment has been used in various fields. Compared with lower resolution infrared images, small targets in higher resolution infrared images have larger average target sizes and occupy more pixel points. Therefore, each pixel in the target has less impact on the pixel-level metrics. A method with good pixel-level metrics does not necessarily indicate a strong target detection capability. We aim to enhance the target-level detection performance of high-resolution infrared small target images through the following series of initiatives. First, we construct a single-frame high-resolution infrared small target dataset. Then, we propose an infrared small target pyramid self-attention network (ISPANet) according to the features of small targets in high-resolution infrared images. Finally, we propose a new loss function Focal Soft-IoU (F-SIoU). F-SIoU loss makes the network more concerned about the hard positive examples. Comparative experiments with state-of-the-art networks have shown that our proposed ISPANet has much better target detection performance in high-resolution infrared small target images. Especially in suppressing false alarms of targets, it has a significant effect, with a 2% increase in target detection probability (Pd) and a 4% increase in target accuracy (Pa). ISPANet effectively enhances the credibility of target detection results. At the same time, we also use the actual acquired dataset to test the ISPANet and obtain good results. The dataset and network will subsequently be made public on GitHub.

Keywords

• Deep learning
• infrared image
• self-attention
• small target detection