Research On Interactive Image Segmentation Method Based On Deep Learning

Interactive image segmentation is to provide useful prior knowledge to computer,and the interested areas are separated from complex environment by user interaction assistant computer.The interactive segmentation method solves the shortcomings of automatic segmentation to a large extent.For many complex scenes in life,interactive image segmentation can accurately separate the foreground and background of the scene,such as adding manual annotation in the target area or manually modifying when the result of automatic segmentation is not ideal.Since the excellent performance of deep learning method in Imagenet competition,the superior performance of deep learning has attracted much attention.Researchers began to explore the application of deep learning in the field of interactive image segmentation.A large number of experimental data and experimental results show that the interactive segmentation method combined with deep learning can not only greatly reduce the user's interactive information,but also save the labor cost and time cost,and the segmentation result is more accurate than the existing methods.After continuous research on interactive image segmentation method based on deep learning,segmentation technology is more and more mature,but there are also the following problems:(1)The existing interactive segmentation method needs a lot of annotation images to train,so it needs high labor cost and time cost.(2)The boundary frame method is not ideal for the image with complex texture and noise,and the result of separating the target area from the complex image is not ideal.(3)The number of pixels that users mark is small and the segmentation efficiency and segmentation performance are relatively good based on the punctuation method.However,for some images,the segmentation results are incomplete,and the segmentation results will have problems such as holes and discontinuities of boundary.In view of the above problems,this paper makes improvement from the following aspects:1.A deep interactive image segmentation method integrating extremum features is proposed.By labeling the extremum points of the target,the interactive information of the extremum points and RGB channel information are fused together to get richer target information and better segmentation results.2.Develop an image annotation software which integrates the extremum features of interactive segmentation method,Compared with the existing segmentation software,this software saves the cost of manual labeling and improves the segmentation accuracy.3.A full convolution dual stream fusion network for interactive image segmentation is proposed.Two independent streaming networks are used to extract depth features from image and user interaction respectively,and the fusion network is used to fuse the two streaming features to predict the foreground and background.Because the two stream fusion structure reduces the number of layers between user interaction features and network output,it can improve the impact of user interaction on the prediction results,so as to obtain better segmentation performance.