Single-Frame Infrared Small Target Detection Network Based on Multibranch Feature Aggregation

Single-frame infrared small target detection is critical in fields, such as remote sensing, aerospace, and ecological monitoring. Enhancing both the accuracy and speed of this detection process can substantially improve the overall performance of infrared target detection and tracking. While deep learning-based methods have shown promising results in general detection tasks, increasing network depth vertically to improve feature extraction often results in the loss of small targets. To address this challenge, we propose a network framework based on multibranch feature aggregation, which expands the network depth horizontally. The parallel auxiliary branches are carefully designed to provide the main branch with semantic information at varying depths and scales. Furthermore, we introduce a differential correction module that corrects erroneous target features through differential methods, significantly boosting detection accuracy. Lastly, we develop a joint attention module that combines channel and spatial attention mechanisms, enabling the network to accurately localize and reconstruct small targets. Extensive experiments on the NUDT-SIRST, SIRST, and NUST-SIRST datasets demonstrate the clear superiority of our approach over other state-of-the-art infrared small target detection methods.