Distortion information and edge features guided network for real-world image restoration

Image restoration methods for specific single distortion have exhibited impressive performance. In real scenarios, distortions are hybrid and variable, and these methods are no longer effective. Recently, some methods for hybrid distortion have been explored. However, their performance drops sharply when the distortion type changes. In addition, the images restored by these methods lack edge details. To solve these issues, we try to learn the distortion information and edge features of the image and use them to guide the reconstruction of the image. Based on this, we propose a Distortion Information and Edge Features Guided Network (DIEFGN). We define a distortion vector to represent the distortion information of the image and use neural networks to estimate it. Since the edges of the image are anisotropic and the 3 × 3 convolution is isotropic, we propose multi-direction linear depthwise convolution (MLDC) to better extract edge features. During image reconstruction, we propose a multi-level progressive fusion strategy to fuse edge features into original image features to enhance the edge details of the restored image. Additionally, distortion vectors are used to modulate the fused image features at all levels, enabling the network to adapt to the variable hybrid distortion. Experiments indicate that the proposed DIEFGN achieves state-of-the-art performance when dealing with real-world images with different distortion types and distortion levels.