Research Of Image Segmentation Of Human Facial Features And Sharpness Matching Method

With the continuous growth of machine learning theory and the rapid development of deep learning technology,the task of human facial image segmentation has come to the attention of researchers in recent years.The facial image segmentation technology is able to realize fine-grained analysis of human faces,and has wide application space in many areas such as facial expression recognition,facial editing,beauty simulation,etc.However,in practice,there are still shortcomings in facial image segmentation techniques.For example,current mainstream facial segmentation techniques use deep learning multi-network fusion models,while high accuracy of segmentation is obtained,the vast structure of networks require expensive equipment and is time-consuming.In additional,in the subsequent facial parts blending applications,the inconsistency of the sharpness between images can also seriously affect the simulation results.Aiming at the above problems,the main research work of the paper is as follows:(1)A facial coarse localization method was proposed based on the facial landmark detection algorithm.Firstly,common facial landmark detection algorithms were compared and analyzed,among which the cascade regression tree algorithm was chosen for implementation.Trained with Helen dataset,the algorithm was able to localize 194 landmarks of each human face with low time cost and high robustness.Furthermore,based on the relation of facial landmarks,a coarse localization method of facial parts was proposed and the appropriate parameters were determined through experiments.(2)Improve the accuracy of facial segmentation based on U-Net full convolutional network,and further improve computing performance by optimizing the network structure.Aiming at inaccurately localized parts such as eyebrows and mouth,common full convolutional networks were analyzed,among which the U-Net model was chosen for training and implementation for higher accuracy of segmentation.Further,based on the sizes of image features and input matrices,the network structure was reduced in complexity,and the computing performance was further improved while ensuring the segmentation accuracy.(3)A method of sharpness matching between images was proposed.Firstly,four common method of no-reference image sharpness evaluation were implemented,as a multi-angle quantitative evaluation tool for sharpness.Then sixteen image blur algorithms of different parameters were implemented as a multi-scale adjustment tool for sharpness.In order to choose an appropriate blur algorithm according to relative sharpness of images,so that the difference of sharpness was minimized,the random forest model was implemented as the core selector for sharpness matching task.Finally,the Poisson fusion algorithm for seamless blending of face template and the target was implemented.