Research On Image Segmentation Algorithm Based On Weak Supervision

Image segmentation technology as one of the four basic tasks of computer vision,with the rapid growth of computing power and data magnitude in recent years,has been widely used in intelligent medical,autonomous driving,intelligent transportation,intelligent agriculture and other fields,and has made outstanding achievements.Behind these great achievements are massive data annotation.For image segmentation tasks,a large number of accurate pixel-level labels are needed to achieve good results in real scenes.However,pixel-level labeling requires a huge amount of work,and obtaining highprecision pixel-level labels is time-consuming and laborious,which greatly limits the scale of the current image segmentation data set.Thus,it hinders the large-scale development of image segmentation technology in different fields.In recent years,many scholars have tried to achieve image segmentation from the direction of reducing the use of pixel-level labels,which is referred to as weakly supervised image segmentation.In recent years,weak supervised image segmentation mainly uses rough but easy to get weak tags such as boundary frame,doodle and image level for image segmentation algorithm training.This effectively alleviates the dependence on pixel level annotation and promotes the development of image segmentation task.Due to the use of rough and inaccurate weak tags,its segmentation accuracy is usually lower than that of the fully supervised image segmentation method.However,due to its lower labeling cost,more and more scholars have been attracted to carry out related research.This paper will focus on the weak supervised image segmentation technology using boundary frame annotation,and put forward two successive methods to improve the performance of the current weak supervised image segmentation method,the specific work is as follows:(1)Because boundary box labels can provide position information that completely contains the foreground object region,in order to make use of continuous background information brought by boundary box labels,a high-universality,plug-and-play weak supervision component Polyp Box is proposed,which can transform the existing fully supervised polyp segmentation method into a polyp segmentation method that only uses boundary box labeling.The module consists of mask projection loss,pixel representation module,fore-background search loss and neighborhood pixel consistency loss.First,the pixel representation module is designed to learn the feature representation of each pixel from the feature map.According to the location information of the boundary frame,multiple prototypes of the fore-background are clustered by KMeans.Then,the forebackground search loss is proposed to search and match pixels in the frame with the forebackground prototype to establish constraints.The mask projection loss constraint model was designed to predict the position of foreground polyps inside the boundary frame.Finally,it was proposed that the neighborhood pixel consistency loss would make the polyp prediction results consistent among pixel point pairs with similar neighborhood.Finally,in the polyp segmentation task of medical images,the feasibility of weakly supervised image segmentation method based on boundary box label was verified for the first time,and the segmentation performance of four polyp segmentation data sets including CVC-300,CVC-Clinic,ETIS,and Kvasir was equal to that of the current mainstream fully supervised polyp segmentation methods.The Mean Dice reached 0.762,0.752,0.474 and 0.810.Moreover,it outperforms the existing weakly supervised image segmentation method on the universal dataset COCO,and the AP reaches 33.2.(2)In order to solve the problem that the representation of part of the Embedding cannot distinguish the fore-background well and it is easy to generate excessive segmentation,a weakly supervised image segmentation method based on transformation consistency and background highlighting erasable is proposed.The method consists of three modules: transformation consistency constraint,receptive field module and background erasure.The transformation consistency constraint enables better differentiation between polyp and background pixel Embedding.The receptive field module extracts the features of polyp region from the multi-scale perspective.Background erasing module improves the discrimination ability of background region and localization ability of polyp boundary by strengthening the response value of background features in the image.Experimental results show that this method can greatly improve the accuracy of previous methods and has better precision for boundary segmentation.On Kvasir data set,Mean Dice is improved by 5% compared with the previously proposed method.