Research On Pixel-level Image Segmentation Based On Generative Adversarial Network

Image segmentation is a primary research orientation in the field of digital image processing.It acts as the basis of many computer vision tasks.Its goal is to classify pixels in images and extract valuable information.With the development of the era,many fields have put forward higher precision requirements for detail segmentation,such as microscopic tumor segmentation in the field of modern medicine.The deep learning image segmentation algorithm can automatically extract the low-level and high-level semantic information of the images,which makes the accuracy of the image segmentation algorithm qualitatively improved,but there still remain the problems such as the loss of edge detail and low boundary segmentation accuracy under complex background.Aiming at these defects,this thesis investigates a variety of deep learning image segmentation methods,on the basis of Generative Adversarial Network(GAN),the research of image segmentation algorithms is carried out.This thesis takes person segmentation as an example,using GAN and soft segmentation technology so that the image segmentation accuracy to achieve more accurate pixel level segmentation.First of all,this thesis studies and compares the recent relatively excellent image semantic segmentation algorithms,such as PSP-Net,Bise-Net,GCN,U-Net,etc.Although these algorithms can complete the task of person segmentation,there is still a problem of missing details of person edge segmentation.In order to further improve the accuracy of pixel-level person segmentation,this thesis introduces the idea of GAN,designs and implements a person segmentation network based on GAN.The U-Net structural generator network is set up for person segmentation,and the high-level and low-level semantic information is connected through the skip connection.The Discriminator distinguishes the generated person segmentation map from the real one.At the same time,the loss function is constructed.It also corrects the inconsistency between the real segmentation image and the generated segmentation image.As a result,the edge segmentation precision under complex background is improved.Secondly,in order to solve the minor problem of the loss of the smaller edge detail such as hair,this thesis introduces a soft segmentation algorithm,which is based on the abovementioned image segmentation algorithm.The boundary-attention portrait segmentation algorithm is designed and implemented,to further optimize the segmentation accuracy of the hair.The main process of the algorithm is divided into four stages.The first stage is obtaining the foreground and background of the portrait through the GAN segmentation algorithm.The second stage obtains the most likely area of the hair through the boundary attention module.The third stage separately feeds the trimap and the original image into a convolutional neural network of encoder-decoder structure for optimal learning to obtain edge detail information.The fourth stage fuses the foreground and edge details to obtain the final subtle segmentation result.Finally,the experiments are respectively conducted on Supervise.ly person dataset and Baidu person dataset,to verify the effectiveness of the segmentation algorithm based on the GAN.Compared with the common segmentation network such as U-Net,after adding the adversarial mode training,the m IOU has increased by 4.56%.The portrait images randomly downloaded from the Internet is applied in experiments,and the segmentation results are better,which proves the generalization ability and application value of the algorithm.The encoder-decoder portrait segmentation algorithm is tested on the portrait segmentation dataset published by Shen.The experimental results show that more accurate detailed segmentation results can be obtained through the optimization of four stages.