Research On Image And Video Segmentation Based On Manifold Learning

Image segmentation is widely applied in image editing,target tracking and other fields,and is one of the most important tasks in computer vision.The concept of fully automatic image segmentation is ambiguous because the part of interest for people is uncertain,and the full manual segmentation takes too much time.Therefore,to achieve the segmentation of the foreground and the background accurately,it is necessary to specify the foreground of interest to the user in some way.Currently,there are two ways to specify users to be interested:(1)The foreground and background are specified in a human-computer interaction manner,and the algorithm of image segmentation in this way is the interactive image segmentation algorithm.(2)Taking the common foreground of multiple pictures as the foreground of interest to the user,the algorithm for image segmentation by this method is the cosegmentation.The existing interactive image segmentation and cosegmentation algorithms are mainly based on the generative model.The algorithm based on the generative model has the advantage of high precision but the modeling process is more complicated,and the discriminant model is simpler and relatively easier to formulate.In addition,when adding new categories,the discriminant model does not need to be retrained.Therefore,it is necessary to design an image segmentation algorithm based on the discriminant model.This paper aims to design an interactive image segmentation algorithm and a cosegmentation algorithm based on discriminant model and study the properties of the algorithm.The main work is as follows:(1)Aiming at the problem of interactive image segmentation,this paper trains the segmentation classifier based on semi-supervised manifold regularization framework to realize the interactive image segmentation algorithm based on discriminant model.The specific method is to transform interactive image segmentation into semi-supervised binary classification.The super-pixel samples are marked by human interaction.Then the semi-supervised learning framework based on manifold regularization is used to train the classifier.Finally,the classifier is used to predict unlabeled super-pixel samples to get the segmentation result.In order to verify the performance of the algorithm and study the nature of the algorithm,this paper conducts experiments on the grabcut and BSD300 datasets.Experiments show that the interactive segmentation method based on manifold learning achieve a good performance and the segmentation error rate on the Grabcut image dataset is smaller than the existing method.The experiment can also be seen that the number of seed points of user interaction has little effect on the segmentation performance,but the number of error seed points of user interaction will greatly affect the segmentation effect.Finally,by observing the segmentation effect of different images.It can be seen that the algorithm mentioned in this paper has a poor effect on image segmentation with large overlap between foreground and background feature space.(2)For the problem of cosegmentation,this paper proposes a cosegmentation algorithm based on multi-task manifold learning.The main steps of the algorithm are obtaining the foreground and background regions of all the pictures by using the cosaliency detection algorithm at first,to mark the super pixel samples,and then add the manifold regularization term to constrain the multi-task learning framework.Classifiers respectively predict the labels of all the superpixels of the picture,and finally obtain the segmentation result.This paper experiments on the iCoseg dataset,and verifies that the algorithm has good segmentation accuracy and extends it to the Jumpcut video dataset,which still has a good effect on the segmentation of video frames.