Research On Interactive Image Segmentation Method Based On Weakly Supervised Learning

With the continuous development of deep learning,automatic image segmentation methods using pixel-level labels to divide any region of interest emerge one after another,and the accuracy is getting higher,even reaching the average level of human recognition.However,pixel-level labels require each pixel to have a corresponding label or background,which requires a lot of labor of labeling personnel and financial costs.In particular,pixel-level labels in the medical image field need to mark areas of lesions,surgical tool areas,etc.,which requires additional domain expertise costs.Therefore,pixel-level labels are difficult to obtain.In this paper,quit a few research and analysis are made on the accurate and expensive pixel-level labels needed by automatic image segmentation technology.It is found that the main methods to alleviate this problem are weakly supervised semantic segmentation and interactive image segmentation.Therefore,this paper makes a further study on the two methods and finds the shortcomings of the above methods: the segmentation accuracy(MIOU)of the former is far from that of the automatic segmentation method using pixel-level labels;Although the segmentation accuracy(MIOU)of the latter is very high,even better than the former,the inference model is obtained by using pixel-level label training.In other words,neither of them is free of costly pixel-level labels.To reduce the reliance on pixel-level labels,this paper decides to combine the features of the former without using pixel-level labels and the latter with high accuracy.The main contributions and contents are as follows:(1)For common image data sets,this paper proposes an interactive image segmentation method based on weakly supervised learning.First of all,in order to get rid of the influence of pixel-level labels,this paper proposes the method of class-activated mapping to process the image data set.In order to better retain the location information and category information,this paper proposes the use of dual-branch structure to purify the information and further make pseudo-labels.Then entered the stage of interactive image segmentation,the use of Pseudo-Label on last step and click on the interactive strategy training in the framework of interactive image segmentation in the interactive inference model,put forward on the basis of the inference model putting to use interaction information of weaky supervision to optimize the segmentation results.Experimental results show that the results are better than other weakly supervised semantic segmentation and close to fully supervised semantic segmentation.(2)This paper proposes an interactive medical image segmentation method based on weakly supervised learning for fine-grained medical image data sets.First,this paper only uses the classification information of medical image data set,and also uses the method of class-activated mapping to retain the location information and classification information of images.Then,in the double branch structure,morphology method is proposed to remove the burrs to purify the location information and category information,and further make pseudo labels.In the stage of interactive image segmentation,an attention mechanism segmentation framework is proposed for fine-grained information,which solves the problem of similar but difficult to distinguish regions.Finally,the interactive inference model is obtained by using pseudo label and interactive click strategy,and user interaction information is introduced to optimize the segmentation result,which is close to the automatic image segmentation method.