Performance Comparison of Irregular Face Inpainting via Deep Learning

As a specific application of image inpainting, face inpainting is critical content in the computer vision. It plays an important role object removal, photo editing and other fields. Deep learning has become mainstream approach inpainting. In applications, corrupted area images usually irregular. For classical irregular approaches based on deep learning, this paper divides it into convolution operator optimization methods structural information constraint methods, former includes PConv GConv latter EC, PRVS, MED, CTSDG. We fist describe basic principle each algorithm detail about strengths limitations. Then we experiment CelebA-HQ dataset, evaluate compare performance quantitatively qualitatively.