Abstract: In order to overcome the problems of edge blur and ghost shadow in image inpainting, a novel inpainting algorithm is proposed in this paper. The algorithm integrates both multi-scale residual blocks and gated convolution operation. The architecture consists of an edge generation network and a texture repair network. Firstly, the algorithm utilizes both multi-layer convolution and multi-scale residual blocks to extract the edges which guide the completion of the damaged area. Secondly, generated edge image is treated as a structural priority and gated convolution is used to eliminate the interference of invalid pixels on updated mask. This results in final image with fine local texture. Experimental results on two benchmark datasets, namely Places2 and CelebA, indicate the proposed algorithm surpasses existing mainstream algorithms in all indicators. It provides PSNR improvement of 2.1%~2.6%, SSIM improvement of 1.6%~2.3% and L₁ loss reduction of 6.5%~9.9% totally.