Research On The Method Of Portrait Beautification For Commercial Photography

With the tremendous development of the computer industry and the immensely improvement of the quality of various image acquisition equipment,portrait beautification in commercial photography has received widespread attention.At present,the post-processing of portrait photography remains handcrafting,which is time consuming and labor intensive.Therefore,it is of great theoretical significance and application value to develop an automatic post-processing software for commercial portrait photography.Most of the existing portrait beautification software is oriented towards images taken by the mobile terminal,which may lead to oversmooth skin,blurred texture and damaged background in color and details when it is applied to photography images.In this thesis,harnessing the power of deep learning technology,three key technologies in portrait beautification are studied,including skin segmentation,wrinkle removal and portrait skin soften.The specific research contents are as follows:Firstly,aiming at the shortage of existing skin segmentation data and the susceptibility to light and skin color,a data optimization algorithm based on pseudo-label learning and an edge guided skin segmentation algorithm are proposed respectively.We refine the pseudo-labels of skin segmentation,which are combined from low-quality human parsing dataset,to high-quality ones.Then,an additional decoder is added to the fully convolution neural network Deep Lab-v3+ to conduct multi task learning by regressing the skin edge contour.The edge information is considered as an attention mechanism to improve the accuracy of skin segmentation task.The experimental results show that the proposed skin segmentation algorithm attain decent performance,and is superior to the existing skin segmentation algorithms in both quantitative test and qualitative test.Secondly,considering the lack of paired data in wrinkle removal task,we propose a facial wrinkle removal algorithm based on CycleGAN.Meanwhile,the directly using of CycleGAN will lead to undesired changes in the generated image in this task.In this thesis,we disentangle the paired generators to alleviate the cooperative cheeting between them,making the loss function guide the training process of the specific generator accurately.Then,a differentiable data augumentation method which can be used in generative adversarial networks is introduced in CycleGAN to stabilize the training process.The experimental results show that the proposed method can greatly restrict the change in the undesired region,leading the precise removal of wrinkles.Thirdly,to alter the eminent drawback of detail missing in the existing algorithm of portrait skin soften,a dual-branch U-Net-like deep convolution neural network model is proposed by fusing the idea of handcraft skin soften technique to individually deal with the high-frequency and low-frequency components.Experimental results show that the proposed method can conduct high-quality portrait skin soften with limited parameters and GPU memory.Based on the above research content,we design and implement a portait beautification system for beautification system.The system encapsulates the skin segmentation,wrinkle removal and portrait skin soften proposed in this thesis into an easy-to-use executable software,and the test verifies that the various functions of the system are operating normally.