MCDF-Net: Dynamic Adaptive Network Based on Modal Competition and Dual Encoder Feature Fusion for Remote Sensing Image Target Detection

Remote sensing image object detection, leveraging the complementary characteristics of infrared and RGB imaging, represents an effective approach for achieving all-weather detection. However, in complex environments, the quality of information provided by different modalities can undergo dynamic variations, necessitating dynamic adjustment of the weights assigned to each modality. Therefore, a Dynamic Adaptive Network based on Modal Competition and Dual-Encoder Feature Fusion (MCDF-Net) is proposed to implement precise modeling of dynamic complementary relationships and adaptive extraction of discriminative features through hierarchical feature dynamic interaction and an adaptive salient modal competition mechanism. Specifically, a Hierarchical Feature Attention Fusion Module (HFAM) is designed under dual parallel feature encoding branches to enable the fusion of global context and local details, in which the Cross-Channel Attention Module (CCAM) is adopted to enhance channel responses through reconstruction via channel feature correlation matrices, and the Difference Fusion Attention Module (DFAM) concurrently calibrates spatial biases through pixel-level difference modeling. Moreover, an Information Entropy-Guided Adaptive Modal Competition Mechanism (IEAMC) is proposed to filter high-confidence queries by quantifying feature point uncertainty, thereby providing useful prior information for the decoder and adaptively determining the salient modality for targets to balance modal contributions. Experimental results over two benchmark datasets, i.e., DroneVehicle and VEDAI datasets, demonstrate that the proposed method clearly outperform state-of-the-art algorithms by effectively handling highly dynamic feature variations.