Hir-net: a simple and effective heterogeneous image restoration network

Image restoration refers to restoring the original image as much as possible from the damaged or degraded image. In recent years, deep learning-based methods have become the mainstream method for image restoration. However, the existing deep learning-based image restoration networks have the same encoder and decoder structures. The homogeneous image recovery network has limited feature representation capability, which limits its image recovery capability. In addition, the homogeneous network does not exploit the relationship between global and local features of the image well, resulting in poor quality of the recovered images. To address the above issues, we propose a heterogeneous image restoration network (HIR-Net). HIR-Net uses Transformer as an encoder to extract the global features of the image, and a convolutional neural network -based feature enhancement block is designed as a decoder to recover the local details of the image. In addition, HIR-Net introduces a new feature fusion strategy to enhance the feature representation capability of the network. This strategy performs cross-attention operations on global features extracted by the encoder and local features extracted by the decoder to achieve cross-fusion of global and local features. Compared with state-of-the-art methods, the proposed method can achieve better performance on image denoising, image draining, and underwater image enhancement.