Infrared dim and small targets detection via self-attention mechanism and pipeline correlator

Abstract

Detecting dim and small infrared targets (DSIT) is always the most concerned and challenging problem in the field of infrared target detection, due to the small size and lacking of high-level semantic information. Benefiting from the great success of deep learning methods in other fields, many superior methods have been proposed in the field of infrared target detection. Based on the analysis of the distributions of a large number of DSIT, we explore to detect DSIT through heatmaps which is more competitive in terms of computational cost and detection performance than using anchor boxes or segmentation maps. Following this line, we design a novel DSIT detection method while adopts the semantic-level and the pixel-level self-attention mechanisms, and improve the Adaptive Pipeline Correlator (APC) in the multi-frame filtering tracking. We evaluate the proposed method on the single-frame database for Infrared Small Target detection (SIRST) and sequence-frame database for Dim Small Infrared target tracking (DIRST). The results show that our method achieves the recall rate and the accuracy rate higher than 98% on SIRST. It takes about 1/9 of the computational cost of the Single Shot MultiBox Detector for Small Target (SSDST) series methods and achieves better results than them on DIRST. When cascading with APC, it achieves an excellent performance with the recall rate of 92.8% and the accuracy rate of 97.9%. All these results fully prove that the proposed detection method is competitive for detecting DSIT.