Co-attentional cascaded perception network for multimodal remote sensing target detection

To address issues such as feature misalignment and additional feature uncertainty caused by heterogeneity, asymmetry, and information redundancy between modalities in multimodal remote sensing target detection technology. Therefore, a multimodal remote sensing target detection model based on a Collaborative Attention Cascaded Perception Network (CACPNet) is proposed, with the aim of improving remote sensing target detection performance. Firstly, the Multi-scale Perception Cross-modal Fusion module (MPCF) is designed for complementary fusion learning. It effectively addresses feature misalignment and enhances training robustness by leveraging Dense Residual Learning Network (DRLN) and Cross-scale Perceptual Feature Module (CPFM). Secondly, innovative progressive attention Hierarchical Cascade architecture is introduced. It collaborates with the Multi-directional Position Coordinate Attention (MPCA) module to reduce information redundancy and suppress background noise for feature quality optimization. This architecture fully leverages the strengths of both the cross-modal fusion module and the feature optimization module, balancing computational complexity while improving detection accuracy. Quantitative and qualitative experiments on the VEDAI dataset demonstrate that our model achieves 83.0% mAP50 and 59.3% mAP50-95. Furthermore, generalization studies on the Drone Vehicle dataset effectively validate the high-quality multimodal remote sensing detection capabilities of our approach.