Efficient feature injection model for visible and infrared images fusion

Visible and Infrared images possess natural complementary characteristics. Their fusion images integrate the strengths of both modalities, extracting richer visual features and significantly enhancing the accuracy of advanced semantic tasks in image processing. To tackle the challenge of cross-modal feature injection, we propose an efficient feature injection model for visible and infrared image fusion (VIF) : EFIFuse. Specifically, we designed a multi-scale feature injection block based on gradient residuals (MFR) and a shared feature injection block based on efficient attention mechanism (SFAM). The MFR leverages multi-scale convolution to capture features across various receptive fields, utilizes the gradient residual module to transmit gradient information effectively, and retains more high-frequency texture details. The SFAM utilizes attention mechanism to extract shared background features between image pairs, facilitating effective information exchange across multi-modal images. We performed experiments on various datasets to evaluate EFIFuse against several state-of-the-art approaches. The results show that EFIFuse outperforms in both measurable indicators and visual appeal.