Research on image restoration combining lightweight transformer and dense feature inference

Limited by the nature of traditional convolutional neural networks such as local prior and parameter sharing, the ability of image restoration to cope with complex semantic environments and large-scale hole repair needs to be improved. In recent years, transformer network architectures based on self-attentive mechanisms have performed well in NLP and high-resolution visual tasks. Compared with traditional CNNs, it is more effective in long-range feature migration applications, but there is high computational complexity in directly using the self-attention mechanism in high-dimensional image tasks. To address this problem, this work improves on the self-attention mechanism by using a linear attention mechanism with gating, a dense feature reasoning module (DFR) embedded in the middle of the U-Net style network, and feature fusion of different codec layers through jump connections. Comparative experiments on publicly available datasets (Paris Street View, CelebA-HQ, Places2), qualitative analyses and quantitative evaluations demonstrate the ability of the approach to reduce computational complexity and to perform well for the repair of broken images with complex semantics of largescale holes.