ECNFP: Edge-constrained network using a feature pyramid for image inpainting

Image inpainting is widely used in the image restoration field. For large-region damaged images, many of the existing methods cannot reconstruct a reasonable structure, which makes the restored images blurry or the structure chaotic. Moreover, these methods restore images with evident differences in color and texture from the undamaged region for large-region images due to the loss of low-level features when extracting high-level semantic information through continuous convolution. To solve the above problems, an edge-constrained network using a feature pyramid for image inpainting is proposed in this work. This network consists of two subnetworks. In the first subnetwork, a rough recovery result is obtained, and edge information is acquired through this result. The structural information of the repaired image is restricted by adding edge information, such that the image structure remains intact. In the second subnetwork, the feature pyramid module aims to obtain low-, middle-, and high-level semantic information by fusing multiscale features, such that the restored area has enhanced color and texture visual effects. This method is compared with the latest methods on the CelebA and Paris StreetView datasets and exhibits better results in terms of peak signal-to-noise ratio, structural similarity index, mean absolute error, and visual effects after restoration.

Automated summary

In this study, a new image inpainting method using a feature pyramid and an edge-constrained network is proposed. It achieves better results in restoring large-region damaged images, including a reasonable structure and enhanced color and texture visual effects.