DC-Net: A Dual-Channel and Cross-Scale Feature Fusion Infrared Small Target Detection Network

In previous deep learning-based infrared small target detection tasks, the problem of detail loss in pooling and downsampling layers was often overlooked in the feature extraction process, and it was not possible to effectively extract global context and utilize the position information in shallow feature maps and semantic information in deep feature maps of infrared small targets. To solve the above problems, first, we designed dual-channel feature fusion net as the backbone network, which removed pooling layers and used advanced dual-channel feature fusion module as the feature extraction module. Then, at the neck of DC-Net, we designed cross-scale feature fusion net suitable for infrared small targets, which performs feature extraction on the features. The information was fully and effectively fused, and finally, we used sharpening algorithm for processing to further enhance the small target area. The experimental results on the single-frame infrared small target (SIRST) and infrared image sequence (IRIS) datasets show that DC-Net exhibits superior performance in both qualitative and quantitative analysis. The mAP50 obtained in SIRST testing increased by 11.7% to 81.4%, on the basis of YOLOv5s. In IRIS, it was 3.7% higher than YOLOv5s to 96.3%.