Research On BEGAN Face Image Inpainting Algorithm Based On Single And Double Discriminant Nets

In recent years,face recognition technology has developed significantly.In the process of face recognition,since there are generally no constraints when acquiring face images,people's postures,expressions,and lighting conditions are complicated,and sometimes there are occlusions of decorations such as hats and glasses.The performance of the existing face recognition technology is degraded,which seriously hinders the practical process of face recognition technology.Most face completion methods require the original complete sample of the object to be repaired,and most existing methods only extract useful information from a single image or use it to generate content similar to the known region.The content lacks the relevance of local and global information,and the repair results lack fine texture details and cannot achieve satisfactory results.Aiming at these problems,this paper proposes a double-discriminant network face image restoration algorithm based on BEGAN(Boundary Equilibrium Gnenerative Adversarial Networks).The main research work is as follows:(1)The BEGAN's discriminant network model is applied to the face image restoration task,and the structure of the generated network is improved.The generation network is not a single decoder structure,but a codec structure,including a deep encoder and a decoder,which is beneficial to the optimization of the convolution network,accelerates the convergence speed of the model,and can initially realize the repair of the occlusion face image.(2)Based on the global discriminant network,a local discriminant network is added.This paper uses two discriminant networks,which not only improves the authenticity of the content generated by the repaired area,but also ensures the visual consistency of the restored image.(3)In order to further constrain the error between the generated image and the real image,reconstruction loss is added based on the reconstruction error of the generated network,thereby increasing the authenticity of the generated image.Finally,aiming at the authenticity and fairness of face completion image quality evaluation,a subjective and objective combination evaluation method is adopted.Both the objective angle and the subjective angle indicate that the improved completion algorithm has higher repair quality and better repair effect.The feasibility and rationality of the algorithm are further verified.