Infrared and Visible Image Fusion Based on Co-Gradient Edge-Attention Gate Network

In order to better emphasize objects, the existing fusion methods tend to filter out high-luminance information that may obscure objects. However, the excessive filtering and indiscriminate fusion may inevitably eliminate certain meaningful scene information. To address this issue, a novel co-gradient edge-attention gate network for infrared and visible image fusion is proposed in this paper. Firstly, a co-gradient-guided edge-attention mechanism that aggregates gradient information of diverse source images is designed to guide bilateral-based fusion focusing on object edges and texture details. Secondly, a gated fusion module is proposed to assign weights on diverse modal information, which fully considers the contribution differences between diverse modal images. Last but not least, a gated fidelity loss is proposed to flow back the difference between fused images and decoupling saliency maps emphasizing infrared high-contrast objects and visible scene details to the gated fusion module. The loss assists in optimizing the amount of diverse modal information flowing through the gated fusion module to better reflect salient object and scene information. The qualitative and quantitative experiments demonstrate that our method performs superior fusion performance compared to other state-of-the-art methods.