Two-stage Network Image Restoration Based on Hypergraphs Convolution
Renshu Li, Dayong Wang, Jie Lu · 2024
In order to effectively repair the complex background and large irregular missing regions and obtain reasonable structure and fine texture, the image restoration based on hypergraphs convolution is proposed. First, the damaged image is input to the coarse restoration network, which propagates the image contextual information to the deeper layers through jump connections. Rich image detail information can be obtained by this way. Down-sampling is used to extract the edge features of the damaged area, and up-sampling is used to restore the details of the damaged edge region, and at the same time, hybrid dilated convolution is used to increase the information sense field, which can help obtain the detailed texture information. Then, the coarse restoration results are input into the fine restoration network with hypergraphs convolution to capture and learn the hypergraphs structure in the input image, and use the inter-correlation matrix of spatial features to capture the spatial feature structure, improve the structural completeness and enhance the detail granularity. Finally, the fine repair results are fed into a discriminator for discriminative optimization to further optimize the repair results. Experimental simulations are performed on internationally recognized datasets, and the experimental results show that the proposed algorithm can generate reasonable structure and rich texture details when repairing large irregular defects.