LPBFusion: A Lightweight Pyramid Branch Feature Decomposition for Infrared and Visible Power Image Fusion

In case of power equipment condition monitoring, vision-based robots are widely used by taking infrared and visible images separately. However, it is still a challenging problem for edge robot to fuse infrared and visible image in real-time. To address the problem mentioned above, this paper proposes a lightweight pyramid branch feature fusion algorithm which fuses infrared and visible image at a fast rate so as to improve the efficiency of equipment condition detection. A lightweight pyramid encoder with branch features is developed for feature extraction, which gives features at four different scales the ability to possess both shared and unique information. We design a differential aware attention mechanism and applied it into the fusion module by which the complementary features of visible and infrared images are better fused. Experimental results indicate that this approach achieves the fastest running speed and superior fusion performance while maintaining a small model size.