Aiming at the problems that the existing image inpainting methods are easy to generate artifacts and do not conform to the original image semantics in the face of large-scale missing, this paper proposed a progressive image inpainting method based on context features. Firstly, the proposed method used the ResNet18 network to roughly fill the damaged image, then input it into a refined network with a dual-branch structure. The context feature aggregation module obtained the region inside the existing image that was most conducive to repairing the image through multi-scale semantic features. It paid attention to the connection between the missing area and the remaining background area in the transfer network learning filled the missing area by higher resolution, and introduced the CBAM module as the network attention mechanism. It defined the global and local discriminant networks to realize the semantic consistency between the generated image and the background and calculated the adversarial loss. It combined the L₁ loss and the structural similarity loss as the network reconstruction loss and then combined with the adversarial loss as the loss function. Experiments on the Place2 dataset show that the average peak signal-to-noise ratio and the average structural similarity are 27.83 dB and 93.19 %, respectively. Comparing with the four image restoration methods, this method is clearer and more natural than other methods in subjective perception, which is highly consistent with the background semantics. In terms of objective indicators, it selects four commonly evaluation indicators for comparison. In terms of structural similarity the method respectively improves 11.48 %, 6.23 %, 3.24 % and 2.21 % that is more in line with human vision. The ablation experimental results of each module of the improved network also verify the effectiveness of the proposed innovations, indicating that the proposed method is superior to similar algorithms.

[2023-05-06] Accepted Paper /doc/2023.03.0073/v1

[2023-11-05] Printed Article /doc/2023.03.0073/v2