A Complex Image Restoration Method Based on Generating Adversarial Neural Networks
ZhenFeng Liu, Jun Wu · 2019
In view of the existing face image restoration methods with occlusion, it is easy to generate problems such as excessive feature learning in the training process and low similarity between the restored face image and the original image. Proposed a based WGAN - GP (Wasserstein Generative Adversarial Networks- gradient penalt) covered face completion method, this method on the model using a generator, a local discriminator and a global discriminator for training the network, to ensure that the generated images and part of the global connectivity. To avoid the single discriminator, network characteristics will study the status of excessive. At the same time, a similarity measurement method is improved to increase the authenticity of the completed image and the original image when the image is completed. Moreover, the similarity measurement of the completed image was conducted through the perception of hash algorithm Dhash and structural similarity SSIM. According to the experiment, in the mixed data set composed of LFW, MegaFace and FaceScrub, the similarity was improved by 4.8%~5.2% compared with other completion methods, which greatly improved the robustness of the image completion method on the blocked face image