A decoupling detection and classification framework for infrared weak and small-target detection in complex background

Infrared weak and small target detection that aims to determine the position of targets in low signal-to-noise ratio infrared images, has always been an open problem in some application scenarios such as sea rescue and military early warning. However, the performance of existing approaches decrease when the target becomes weak in complex background because the following challenges: (1) the contour of infrared weak and small targets is blurred, and the edge and texture features of the targets are very weak, making it difficult to be represented; (2) The complex background contain many noise points such as such as scattered clouds and waves present that are highly similar to the target, resulting in a high false alarm rate. To cope the above challenges, a decoupling detection and classification framework for the Infrared weak and small target detection (DDCISTD) in complex background is proposed, which contains two parts: (1) a spatial attention transformer-based network is designed to represent the target efficiently based on the position of the target and other regions in the image; (2) the false alarm elimination network (FAENet) for eliminating false alarms based on visual concept learning is learned, enabling the training of the network under conditions where the number of false alarms and target samples is imbalanced. Further, the single-frame infrared weak and small target (SIWST) dataset is constructed to evaluate the performance of the approaches on weak small targets because existing datasets lack weak targets with blurred edges.