Image Structure-Induced Semantic Pyramid Network for Inpainting

The existing deep-learning-based image inpainting algorithms often suffer from local structure disconnections and blurring when dealing with large irregular defective images. To solve these problems, an structure-induced semantic pyramid network for is proposed. model consists of two parts: the edge content-filling network. U-Net-based restores defect residual blocks. map input into together in prior condition. In network, attention transfer module (ATM) designed to reconfigure encoding features each scale step by step, recovered feature linked decoding layer corresponding potential fusion improve global consistency finally obtain restored image. quantitative analysis shows that average L1 loss reduced about 1.14%, peak signal-to-noise ratio (PSNR) improved 3.51, structural similarity (SSIM) 0.163 on CelebA-HQ Places2 datasets compared current mainstream algorithms. qualitative this not only generates semantically sound content as a whole but also better matches human visual perception terms connectivity texture synthesis.