SE-DCGAN: a New Method of Semantic Image Restoration

Image restoration is a technique that utilizes the edge of a corrupted area. The information content of the damaged information area is inferred based on the remaining information of these images, and then the damaged area is filled to achieve image restoration. To solve the problem of image occlusion in practical applications, a squeeze-excitation network-deep convolution generative adversarial network (SE-DCGAN) was proposed. First, many new sharp images are generated using SE-DCGAN. Then, in the generated image, the most similar image is found based on the context semantics of the original image and the encoding of the unfilled portion to fill the original image. SE-DCGAN introduces maxout activation with powerful fitting capabilities to improve image generation efficiency and avoid image generation redundancy. Experiments based on three datasets of CelebA, Street View House Number and anime avatars, showed that our method successfully predicted a large number of missing regions. This method improves the recognition rate of occluded images, produces high-quality perceptual results, and is flexible enough to handle a variety of masks or obstructions.