Infrared small target detection based on the combination of single image super-resolution reconstruction and YOLOX

Abstract

For the infrared search and tracking system, it is necessary to increase the ability to detect small infrared targets against complex backgrounds. YOLOX is a high-performance detector, but its detection performance is constrained when it uses data from low-resolution infrared images with small targets. However, occasionally design constraints and budgetary restraints will prevent the optical system and sensor resolution from being increased enough to improve image quality. Real-ESRGAN is used to solve this issue by reconstructing a high-resolution infrared image from its low-resolution counterpart, which will be used as YOLOX-S's input. Also, the YOLOX-S training strategy is modified further to make it appropriate for the detection of infrared small targets, including the Mosaic and MixUp data augmentation and the size of ground-truth. The average precision achieved by the suggested method in this work increases from 63.70% to 77.19%, which shows a considerable improvement in infrared small target detection when compared with the original model by inputting original images.