Auto Portrait Segmentation Based On Deep Learning

Separating the foreground and background in the picture is a long-term research hotspot in machine vision.With the rapid development of image segmentation in security surveillance,autopilot,and photo processing,human image segmentation has become an important issue in image segmentation.However,due to the influence of background changes of the human image,different wearing clothes and various skin color postures,the automatic segmentation of the human image faces a very big challenge.At present,the commonly used human image segmentation methods are based on traditional and deep learning-based segmentation methods.For the traditional image segmentation method,the user often needs to outline the lines in the foreground to assist the segmentation,which cannot be automated.This results in low efficiency and labor cost when it is necessary to process a large number of pictures,and the segmentation effect is often unsatisfactory when the picture is complicated.The method based on deep learning aims to realize general image segmentation,and does not realize high-precision human segmentation effect,and the segmentation boundary is blurred.Therefore,how to reduce labor cost and improve segmentation accuracy is the research focus of human image segmentation.In this paper,the image segmentation and deep learning techniques are studied,and the method of automatic segmentation of human and edge refinement is proposed.The main work and achievements are as follows:1.An automatic segmentation network for human images is proposed.This paper improves on the basis of fully convolutional network,introduces deep separable convolution and improved activation function,which improves network recognition accuracy and speed;adds human body location recognition in the network to assist network segmentation,and perform filter optimization on segmentation results.Finally,the network is verified on the collected image data of the person,which proves the effectiveness of the proposed automatic segmentation network of the person image.2.An improved image edge segmentation method is proposed.In the automatic segmentation network of the human image,the edge region of the segmentation result is not fine enough,so additional edge segmentation processing needs to be added at the back end of the network.In this paper,the edge segmentation algorithm is studied.On the basis of the closed-form solution algorithm,the global color space model is added,which solve the problem that the original algorithm only has local color space and the result is inaccurate.The dynamic local optimization window is added to solve the closed-form solution.The algorithm fixes the problem caused by the window,thus optimizing the edge segmentation effect of the human.Finally,experiments on the collected image data of the characters show that the proposed edge segmentation algorithm has an improved effect on the automatic segmentation network of characters.3.After segmenting the foreground of the human,this paper studies the non-photorealistic drawing of the picture and realizes the cartoonization the foreground of the separated person.This paper also proposes a human image automatic processing system,which supports both Windows and mac OS platforms.After the user inputs a human image to the system,the foreground of the automatically segmented human can be obtained without manual interaction,and the cartoon can be optionally processed.Finally,the feasibility of the system was tested and verified,and good results were obtained.