Deep Interactive Image Segmentation Algorithm Based On Extremum Feature

Separating the desired target image from an image is the essence of image segmentation,which provides a basic operation for subsequent research on image processing,which shows the importance of image segmentation in image analysis.In the image segmentation,the image segmentation can be divided into automatic segmentation if the user does not participate in the image segmentation,and user participation in the segmentation,that is,the interactive image segmentation.Interactive image segmentation is proposed because the computer does not fully understand the content of the image and needs human participation to guide the segmentation.It is now widely used in many fields of image processing.The research purpose of interactive image segmentation is to annotate the target area in a simple,efficient and time-saving interactive way,and the annotation information can be used as an effective guide to extract high-level semantic features in images for accurate image segmentation.With the continuous development of deep learning,deep learning has achieved good results in image segmentation,and end-to-end network models have been used by many segmentation algorithms,but in practical applications,interactive image segmentation is still used in many scenarios.For example,in the medical field,the target area of medical images is marked,or the results of automatic image segmentation are not ideal and require manual correction.However,deep learning can exert its superior performance in the field of interactive image segmentation,and use its superiority in image feature extraction to improve the segmentation accuracy.It is widely used in interactive image segmentation,such as medical image processing,unmanned driving,face However,there are a series of problems such as difficulty in labeling and time-consuming,and low segmentation accuracy.Aiming at the cumbersome and low efficiency of traditional interactive image segmentation,the main research contents and innovations of this thesis are as follows:(1)Propose an interactive way that is easy to operate and obtain more image information.The user only needs to annotate the extreme points at the leftmost,rightmost,uppermost and bottommost ends of the target area,and solve the extreme value box and the center point of the target area according to the position characteristics of the extreme value points,and use the four sides of the extreme value box to solve the problem.The symmetrical point out of the center point is the background "marking point".Using the information of extreme points and background "label points" can guide the network to perform image segmentation.This interaction mode can not only save the user’s time,but also can obtain rich image information under the condition of simple operation.(2)Improvement of network structure.The paper takes the residual network as the basic framework,and improves the residual structure on this basis.The idea of pre-activation is added to the residual structure,and then the BN layer is added;the traditional convolution layer is improved,and the context self-calibration convolution module is replaced by the traditional convolution layer;the network structure is simplified,and the last fully connected layer and the fourth and fifth layers are deleted.The pooling layer of the stage is replaced with an atrous convolution module to ensure the resolution of the image.Finally,a pyramid scene parsing module is added to fuse rich contextual feature information.It is experimentally verified that the Iou of the segmentation results guided by extreme points and background "label points" in the improved residual network can reach 94.3%.