An ID Photo Inpainting Algorithm Based On Deep Learning

Preprocessing for ID photo is the first step in many face recognition systems,and image inpainting is the key to preprocessing.This paper presents two fast ID photo inpainting algorithms based on deep learning,the main contents are summarized as follows:Firstly,we design an ID photo inpainting algorithm based on edge detection.The inpainting task is divided into two steps,the first step is to detect the region to be repaired,the second step is to repair the image.A noise detection network is designed for step one,and the noise position is located by multiple side outputs.A noise inpainting network is designed for step two.Combined with the output of the noise detection network,the image to be repaired of any size can be repaired by quickly passing through the noise inpainting network.Secondly,simulating the cause of broken ID photo,an ID photo inpainting algorithm based on feature separation is designed.The noise feature and face feature are generated by encoding network.And the repaired images are generated by the face feature.Using three parallel tasks,namely outputting the noise position,outputting the reconstructed input image,outputting the repaired image,to complete the image inpainting.The tasks of generating reconstructed input image and repaired image share the decoding network so that it can enhance the degree of feature separation.In order to further improve the quality of image inpainting,four optimization items are used,namely,perceptual loss,regression residual,weighted loss and generative adversarial nets.Perceptual loss solves the average problem brought by euclidean-loss.Regression residual preserves more of the original image information.In order to solve the problem of insufficient inpainting caused by sample imbalance,a weighted loss is designed,that is,noise pixels are given a greater weight in calculating loss than normal pixels.Experiments show that the weighted loss makes the noise-polluted pixel has a more reasonable value.Generative adversarial nets makes the repaired images contain more detail information,which is good for subsequent face recognition.Finally,Peak signal-to-noise ratio(PSNR)and the area under the ROC curve(AUC)are used in the experiments to evaluate the effectiveness of the two algorithms and the optimization items.On the collected databases NEU-PSNR and NEU-Veri,the algorithm based on edge detection gets the PSNR 29.89dB and AUC 0.8203,the algorithm based on feature separation gets the PSNR 29.66dB and AUC 0.8239,which are much higher than the PSNR 19.29dB and AUC 0.6181 obtained by corrupted image.The algorithm proposed in this paper has a good inpainting effect on a variety kind of noise and can run in real time(80fps).