SFCANet: Channel Attention in Spatial-Frequency Domain for Infrared Small Target Detection

Recently, the field of infrared small target detection in the spatial domain has seen rapid development. Nonetheless, distinguishing noise that closely mimics the target in the spatial domain remains a formidable task when relying solely on multi-scale spatial features. Consequently, it is of great significance to explore the combination of frequency and spatial domain characteristics to aid in the discriminative process. Building on this, we propose the Spatial-Frequency Channel Attention Network (SFCANet), which is composed of the spatial-frequency channel attention module (SFCA) and the deep supervised multi-task ensemble learning module (SMTEL). By fusing multi-scale spatial and frequency features, the SFCA refines the process of target feature extraction, thereby enhancing the capability for continuous modeling complex backgrounds. This aids in the discrimination between noise in the background and actual faint small targets. Furthermore, we introduce SMTEL to mitigate information loss in deep supervision multi-task learning, particularly during extensive up-sampling processes at low resolutions. Our SFCANet, by integrating multi-scale spatial and frequency domain information, effectively directs the attention of network to the continuous modeling of backgrounds, while also compensating for the information loss caused by aggressive upsampling. This effectively enhances the detection accuracy for small infrared targets. Experiments conducted on three public datasets, IRSTD-1K, NUDT-SIRST and SIRST-V1 demonstrate the superiority of SFCANet in infrared small target detection. Our code will be made public at https://github.com/linzijin1238/SFCANet.