A Study Of Face Completion Using Generative Adversarial Networks

Face completion is an important topic in computer vision and image processing.The core task of face completion is to restore image information and ensure the completion result is consistent with the ground truth.Because the existing face completion methods ignore the symmetry feature of the face images,and doesn’t constrain the consistency between the completion result and ground truth,which makes it impossible to generate a photo-realistic and consistent completion result for any part of the face.In addition,most of the existing completion methods are not suitable for high-resolution face image completion.To solve those problem,we make study on face completion method.The main contributions are as follows:Firstly,in order to ensure that the completion results are photo-realistic and consistent with the ground-truth,we propose a generative face completion method based on generative adversarial network.Among them,a generator which adopts ”u-net”network structure is used to generate completion result,global discriminator and local discriminator optimizes the global semantic structure and local content authenticity of the completion result in the form of adversarial loss respectively.In addition,we also adopt structure loss and pixel loss to constrain the consistency of the completion results with the ground truth.Then,we propose a symmetry-aware face completion method to optimize the symmetric components completion.The optimization of symmetry completion consists of two steps,the first step is symmetric components detection and the second step is symmetric components completion optimization.For symmetric components detection,we propose a heuristic detection method,which can greatly improve the detection accuracy.For symmetric components completion optimization,a symmetry adversarial loss and a symmetry pixel loss are adopted to optimize the detected symmetric components.Finally,we propose a high-resolution face completion method based on the symmetryaware face completion method.we adopt an improved progressive training method to train our high-resolution model and style loss to optimize the high-frequency details of the completion results.The qualitative and quantitative experiments show that our face completion methods have a greater improvement not only in visual effects but also in evaluation indicators compared with existing methods.