"EFD-Net"

"We propose an Evolutionary Fusion-Adaptive Dual-Stream Object Detection Network (EFD-Net), an end-to-end framework that jointly optimizes fusion quality and detection performance. First, the framework utilizes a dual-stream architecture with independent branches to ensure the integrity of modality-specific representations. Specifically, a Strip-wise Feature Enhancement Module (SFEM) and a Collaborative Spatial-Channel High-order Attention (CoHA) block are developed to enable high-order feature interactions across hierarchies without compromising mono-modal feature discriminability. Furthermore, an Embedded Attention Interaction Module (EAIM) introduces parallel connections to capture global complementary information. To resolve the optimization conflicts that cause fusion degradation, an Evolutionary Algorithm (EA) is employed to dynamically search for the optimal balance of loss function parameters during training. Extensive evaluations on the LLVIP and M3FD datasets demonstrate that EFD-Net alleviates fusion degradation."