Dual-channel feature extraction Hybrid Attention Network for Detecting Infrared Small Targets

Abstract For military early warning, forest fire prevention, and maritime search and rescue, infrared small target detection is critical. However, because of the low contrast and inconspicuous features of infrared small targets, rendering most existing methods ineffective in restoring target edge details or misidentifying the background as a target. This paper proposes a dual-channel feature extraction network (DCFE-Net) with hybrid attention, which enables the network to suppress the background and enhance the target by designing dual-channel feature extraction and multi-layer feature fusion. Specifically, the dual-channel mainly consists of a convolutional attention fusion module, which adaptively integrates feature map correlations by introducing a hybrid attention module to capture global information while enhancing the feature representation of small targets, and a feature compression extraction module, which utilizes depth-separable convolutional combinations to carry out fine-grained target feature extraction while reducing the loss of details. In addition, the multilevel feature enhancement module ensures that the network can capture targets at different scales through skip connection operations, while avoiding small targets from being overwhelmed by deep features, making them simultaneously semantically informative and detailed. Therefore, the network can fuse multilevel features for effective information extraction. According to the experimental results, DCFE-Net performs best in false alarm rate and detection probability. In particular, DCFE-Net has good detection performance with a Pd (0.9925), Fa (1.17×10-6), and IoU (0.8929) on the NUDT-SIRST dataset.