Research And Implementation Of Human Matting In Natural Scene

Image matting is always a basic problem in the field of digital image processing.Its basic definition is to separate the significant object from the original image.In image editing,designers often need to manually matting through matting software,which greatly increases the time of image editing.With the development of social network,people are not satisfied with a single way of text communication.Many people choose to share daily pictures to communicate with people.People often need to edit images in daily life.An accurate human image matting system is of great significance to image editors or people who need matting in daily life.In view of the above scenarios,this paper studies the matting algorithm in natural scenes,and applies the matting algorithm based on deep learning to real scenes,which has a wide range of use value.Matting problem is a ill posed problem.To solve this problem,we usually need to use additional prior information.However,prior information needs to be obtained by manual marking in previous studies,which greatly reduces the ease of use of matting algorithm.In order to achieve the goal of automatic matting,through the study of the relationship between the segmentation task and matting problem,the segmentation result of morphological changes in digital image is taken as the additional prior information of matting task,without the step of manual marking.According to the performance comparison of different segmentation algorithms based on deep learning on human image segmentation data set,the segmentation algorithm with the best performance is selected as the front step of matting task.The performance of different matting algorithms in the natural scene matting dataset is studied in depth.Through comparative analysis,it is found that the matting algorithm using depth neural network technology has the best performance in the task of matting.In addition,Pytorch the python deep learning framework and Pyqt5 image interface framework are respectively used to build the neural network model of segmentation and matting and the design of system image interface,further completing the construction of the whole automatic human image matting system.The automatic image matting algorithm is evaluated on the human image matting data set.The experiment shows that the segmentation based matting algorithm in the human image matting data set can get an effective matting effect.At the same time,the real scene human image data is used to evaluate the automatic human image matting system.The experiment shows that the system can meet the matting timeliness and ease of use.