Efficient multispectral infrared dim target detection framework based on slice registration

In recent years, infrared dim-target detection has emerged as a pivotal area of research. However, most existing detection methodologies focus on single-spectral imagery. Owing to the optical diffraction limit, the image resolution obtained from different detection bands under the same conditions varies significantly. Single-spectral infrared imagery offers limited target and background features, failing to capture a comprehensive representation of the environment and thus struggling with target detection in complex backgrounds. Although multispectral image fusion can enhance the detection capability for dim infrared targets, processing across all spectral regions leads to increased computational complexity, resulting in time-consuming and redundant detection algorithms. In response to this challenge, we propose an efficient multispectral infrared dim target detection framework based on slice registration. The framework consists of a reference spectral rapid localization module (RSLM) and a multispectral feature enhancement detection network (MFE-Net). The latter includes a feature extraction module, a multispectral information-weighted fusion module (MIWF), and a detection module. Initially, potential target locations in the image are rapidly identified through the reference spectral rapid localization module, and corresponding image slices are extracted from other spectral bands based on spectral coordinate transformation. Subsequently, the multispectral feature enhancement network's feature extraction module processes these multispectral target slices to extract features from each band. Finally, the MIWF module integrates information from different spectral bands to enhance the network's sensitivity to infrared dim targets, allowing the multispectral feature enhancement detection network to conduct precise detection, reduce false alarms, and improve detection rates. The proposed method utilizes the reference spectral rapid localization module to reduce the complexity of multispectral data fusion, while the multispectral feature enhancement detection network leverages information from different bands to enrich target features, thereby enhancing the accuracy of weak infrared target detection. Experiments conducted on a comprehensive dataset demonstrate that this method outperforms other state-of-the-art methods in terms of detection probability (Pd) and false alarm rate (Fa).