Video Infrared Small Target Detection Combining Hybrid Attention with Cross-Scale Feature Fusion

Infrared small target detection is widely used in marine resource utilization, high-precision navigation, ecological environment monitoring, and other fields. Traditional detection methods rely on manual feature templates and prior knowledge, which are not robust in complex scenarios. Although achieving promising results, existing deep learning-based methods mainly focus on single-frame target detection. They heavily rely on spatial information of small targets, which cannot generalize well on visually non-salient scenarios. This paper aims at fully fusing temporal and spatial information with a newly designed hybrid-attention and cross-layer feature fusion module. Specifically, we take continual infrared small target frames as input and sequentially process the frames with 3D feature extraction and a 3D hybrid attention module. 3D cross-scale and fusion module (3D-CFM) can ensure the interactive fusion between the underlying structural information and the deep high-level semantic information, both in the spatial and temporal dimensions. Based on the 3D hybrid attention module (3D-HAM), we can enhance the response of target features in the deep layer of the network, and thus further increase the response of the target. To prove the effectiveness of our method, we developed a video infrared small target dataset, which contains 19 infrared small target sequences. They are synthesized by real infrared background scenes with various, moving virtual infrared targets. Experimental results demonstrate the superiority of our method on the probability of detection (Pd) and false alarm (Fa).