Research On Image Segmentation Based On Weakly Supervised Learning

Image segmentation is one of the key topics in the field of computer vision.It is widely used in the fields of target tracking,pedestrian detection,traffic monitoring and medical image analysis.However,for the task of strongly supervised image segmentation,collecting large-scale accurate pixel-level labels requires expensive manpower and material resources.At the same time,as the image scene becomes more complex,the accuracy of the annotation will also decrease.The image segmentation method based on weakly supervised learning can achieve reasonable performance by using coarse labels with less supervision information and lower labeling cost,such as bounding box,scribble,point and even image-level labels.This thesis proposes two weakly supervised learning image segmentation methods to solve the problem of semantic segmentation and instance segmentation relying on data labels.The main work is as follows:Firstly,aiming at the problem of weakly-supervised semantic segmentation,a method of weakly-supervised semantic segmentation enhanced by the data crawled from the web is proposed.The method consists of two modules,the first module cleans the data crawled from the web.Use the target dataset to train the class activation maps,which can filter the images whose object region is too small and the image content does not match the category in the web dataset.The second module is responsible for semantic segmentation,and this module consists of two branches.The first branch uses the previously trained class activation maps plus the dense conditional random field to obtain the initial segmentation mask of the target dataset.The second branch is responsible for segmentation of web dataset.Firstly,use Co-segmentation to train the cleaned web dataset to generate the segmentation mask of the web dataset;secondly,train the fully convolutional network of the web data in the form of fully supervision;Finally,input the target dataset into the trained fully convolutional network to generate an enhanced mask.The last step of the module is to fuse the initial mask generated by the first branch with the enhanced mask generated by the second branch to generate the final segmentation mask,and train the fully convolutional network of the target dataset based on the final segmentation mask to obtain the semantic segmentation results.Experiments show that the m Io U values of the method in the PASCAL VOC2012 train dataset and test dataset are 63.2% and 64.5%,which are better than similar weakly-supervised semantic segmentation algorithms and can effectively boosts the semantic segmentation performance.Secondly,aiming at the problem of weakly supervised instance segmentation,a twostage instance segmentation algorithm is proposed.The first stage focuses on generating pseudo segmentation masks.Firstly,training the class activation maps with peak stimulation layer to roughly locate the target.Secondly,obtain the peak response maps through peak back propagation to obtain the approximate range of the target.Finally,the filling module is used to recover the entire object region from the incomplete peak response maps of the object region to obtain the instance activation maps.In the second stage,the instance activation maps is used as the pseudo mask,and the instance segmentation network is trained in the form of fully supervision.Experiments show that this method is not only better than the method using the same supervision information,but also better than the method using bounding box labels.