Research On Low Quality Watermarked Face Inpainting Via Joint Residual Network

Face verification is an important technology in the field of information security,and it has been widely used in airport passport verification,corporate attendance,and mobile payment systems etc.In order to prevent the face image in the system from being stolen or misuse,the application system with higher security level will perform watermark encryption processing on the stored face image.However,the addition of the watermark will pose great difficulty to accurate face verification.Therefore,how to solve watermarked face image inpainting and improve the performance of face verification has become a problem to be studied urgently.The existing image inpainting algorithms are mainly divided into two categories:image inpainting algorithms based on traditional methods and image inpainting algorithms based on deep learning.The former assumes that the position of the repaired area(watermark)of the input image is known and needs to be manually marked,thus failing to meet the need for automated online processing.The latter only needs to know the position of the repaired area(watermark)of the training sample images,and it is not necessary to know the position of the repaired area(watermark)position of the input test images,thereby avoiding manual intervention.However,in practical applications,the watermarked face in the face verification system is compressed low-quality image.Accordingly,the deep learning training sample images must also be compressed low-quality images.However,the quality loss may cause the position of the watermark changes,and the prior information of the wrong watermark location will cause the network to converge in a poor direction,resulting in the fact that watermark lines will be in the face.In view of the above problems,this paper treats the watermark in the face image as a kind of special noise whose location is unknown,and transforms the watermarked face inpainting into an image denoising problem which does not need to know the position of the input images and the sample images?And thus avoiding the problem that watermark misplacement results in the failure of the image inpainting.Based on this consideration,this paper proposes a low-quality watermarked face inpainting algorithm based on joint residual network.The algorithm solves the problem of watermark inpainting from the three aspects of global inpainting network,local inpainting network and discriminant constraint network.First of all,at the global inpainting level,this paper is inspired by the residual learning theory and takes advantage of a global inpainting based on residual learning network to solve the problem that the watermark in the image is not clean and the details are blurry.Secondly,at the local inpainting level,this paper makes full use of the characteristics of human visual perception and the inherent structure of the human face,and proposes a local inpainting network based on face local constraints to solve the problem of the blurry face local details.Finally,at the discriminative constraints level,this paper considers the importance of image high-dimensional features in face verification and proposes a discriminative constraint network to further improve the performance of face verification.By learning the high-dimensional features of the images to constrain the processing of image details,the performance of face verification is highly improved.The experiment demonstrates the superiority of our low-quality watermarked face inpainting algorithm based on joint residual network,compared with the state-of-the-art methods.The average PSNR of the reconstructed face images is increased by 4.16dB,and the average SSIM is increased by 0.08.TPR is improved by 16.96%when FPR is 10%.It's shown that our algorithm is of great significance for the practical application of watermarked face verification.