Image Inpainting Based on Structural Constraint and Multi-Scale Feature Fusion

When repairing masked images based on deep learning, there is usually insufficient representation of multi-level information and inadequate utilization long distance features. To solve the problems, this paper proposes a second-order generative image inpainting model Structural Constraints Multi-scale Feature Fusion (SCMFF). The SCMFF consists two parts: edge repair network network. combines auto-encoder with Dilated Residual Pyramid (DRFPF) module, which improves semantic structural details images, thus achieves better repair. Then, embeds Attention (DMAF) module in for texture synthesis real as prior condition, fine-grained under constraint by aggregating long-distance features different dimensions. Finally, results are used to replace edge, networks fused trained achieve end-to-end from complete image. compared advanced methods datasets including Celeba, Facade Places2. quantitative show that four metrics LPIPS, MAE, PSNR SSIM improved 0.0124-0.0211, 3.787-6.829, 2.934dB-5.730dB 0.034-0.132, respectively. qualitative distribution center hole reconstructed more uniform, effect line human visual perception.