ECAFusion: Infrared and visible image fusion via edge-preserving and cross-modal attention mechanism

Infrared and visible image fusion (IVIF) strives to integrate detailed textures with clear thermal objects, thus obtaining an information-rich image to enhance the performance of downstream tasks. However, existing image fusion networks generally overlook critical information, which includes edge details, and fail to facilitate effective cross-modal information interaction when extracting features across modalities. As a result, the fused outputs often fall short of expectations. To overcome this challenge, we propose ECAFusion, a dual-branch fusion model for IVIF that incorporates an edge-preserving and cross-modal attention mechanism. Specifically, we design a Sobel convolution block (SCB) by applying the Sobel operator to preserve edge details. Additionally, a cross-modal interaction module (CMIM) is proposed to capture complementary information from different modalities. Finally, we employ a well-designed feature fusion module to combine these features from different modalities. Comprehensive experiments show that ECAFusion surpasses the most advanced fusion methods. Moreover, our model significantly improves the performance of high-level vision tasks.