FocusNet: An Infrared and Visible Image Fusion Network Based on Feature Block Separation and Fusion

Infrared and visible image fusion enables the generation of highly readable fused images through complementary imaging characteristics. They ignore the need to consider different processing methods for primary and secondary information in algorithms serving detection tasks. To improve the performance of fusion algorithms in practical tasks, we propose FocusNet, an infrared and visible fusion method based on the separation and fusion of primary and secondary features. First, we perform a segmentation operation on the input image prior to the fusion task to separate the primary and secondary feature blocks. This segmentation step facilitates the targeted and effective processing of each type of information. Second, we design a fusion network that adaptively adjusts its output based on the available feature information. This adaptive tuning allows for better integration of primary and secondary features, resulting in improved fusion results. Finally, we use a multi-task training strategy on the fused network for training primary and secondary feature blocks. This method allows the network to serve two types of feature fusion tasks simultaneously, keeping the fusion network lightweight. Qualitative experiments and object detection tasks performed across diverse datasets demonstrate the superior result readability of FocusNet.