Mechanism-Driven Low-Altitude Small-Target Detection Through Microwave and Infrared Information Fusion

Detecting low-altitude small targets in complex scenes is highly valuable. However, modal differences and registration errors between infrared and microwave data still pose significant challenges to accurate detection. Therefore, this article proposes HDCNet, a mechanism-driven detection method guided by physical knowledge, which integrates microwave and infrared data into a novel cross-modal fusion framework. The method introduces a new feature alignment technique using self-attention mechanisms and a transformer architecture to map microwave and infrared data into a shared semantic space, reducing modal disparities. To address registration errors, a hybrid adaptive contrastive learning and error correction mechanism is proposed, dynamically adjusting learning weights and geometric constraints to improve cross-modal feature alignment. During radar data preprocessing, a Range-Doppler (RD) transformation is applied to convert radar data into RD maps, capturing motion and reflection characteristics to ensure physical consistency in fusion and alignment. To meet real-time and high-accuracy detection requirements, a lightweight detection network is designed with a decoupled detection head and an optimized dual-label assignment strategy, improving the detection of weak-feature targets. Experimental results show that the proposed method reduces semantic gaps between modalities, achieves precise feature alignment, and enhances small-target detection performance. This work offers a robust solution for infrared and microwave information fusion in complex environments.