Interactive Image Segmentation Method Based On Candidate Boundary Points

With the rapid development of multimedia big data and artificial intelligence,how to extract useful data from massive data has become an urgent need for the development of artificial intelligence.Among them,image segmentation is a basic and important research content in the field of multimedia data analysis and understanding.Due to incomplete computer interpretation of image content,interactive image segmentation algorithms with human intervention have been deeply studied in recent years and have been widely used in many fields of image processing.The purpose of interactive image segmentation is to quickly and accurately extract foreground objects with high semantic features from input images by simple and intuitive human annotation.Now deep learning in many application of image segmentation tasks better and better results have been achieved,including most of the segmentation algorithm using a end-to-end network segmentation automatically,however in reality many scene still need to use the interactive image segmentation method,such as in the annotation of radiotherapy target area,or in the automatic segmentation result is not ideal and need the user manual correction,in addition,when training deep learning model usually need to manually tag many image segmentation results as the training set,using efficient interactive segmentation tools to make the image when the training set more convenient labeled images.Considering the superior performance of deep learning,using it to do interactive image segmentation can reduce the number of user interactions and the time spent,so as to get a more efficient interactive segmentation tool.Through the study of interactive image segmentation,it can be found that interactive image segmentation algorithm can be widely used in medical image processing,biometric recognition,unmanned driving and other fields,but there are also a series of problems such as labeling difficulties,so more simple and efficient interactive segmentation method is needed.The main research content and innovation points of this paper include the following aspects:1,put forward a new interactive way in the image preprocessing step,the user only in the center of the image target marked with the target edge two points,and can quickly find out the target of the candidate boundary point accurately,and the existing methods(bounding box,line drawing,mark foreground and background),compared to the method in the case of greatly save mark time precision and existing popular method can also be obtained the same results.2 HeatMap is generated after obtaining the image target candidate boundary points.Specifically,a separate two-dimensional Gaussian center was generated at each candidate boundary point in this paper,and the HEATmap was used as an additional channel in the segmentation network input of this paper,which was connected with the RGB channel of the original image as a 4-channel input.The segmentation network learning in this paper converts this information into the segmentation of target objects matching these candidate boundary points.3 improve existing network framework,this article will ResNet101 as backbone,and on the basis of improving the residual block structure,and remove the network last full connection layer and fourth largest pool of 5 stages to ensure that the image resolution,at the same time joined the empty convolution to keep the receptive field of the same size with the original network,at the end of the paper web use pyramid scene analysis module,the purpose is to the global context information is added in the final figure.