CRFs Based Image Semantic Segmentation

With the rapid development of mobile devices and the Internet,the image quantity has shown an explosion trend.How to effectively manage and analyze large amount of image information is one of the key problems for modern science and technology.Facing an image,human vision system is able to identify various objects and their subtle contours.The intelligent machines are expected to accomplish a similar task through a series of learning,namely image semantic segmentation.As the fundamental topic in computer vision,semantic segmentation is important for many applications ranging from object recognition to image editing.To tackle the limited expression ability of traditional Conditional Random Fields(CRF)model,researchers propose to integrate sparse dictionary learning into CRF.However,the existing solutions merely focus on the discrimination of dictionaries,ignoring the inherent data locality characteristics.In this paper,we propose a novel semantic segmentation method based on an innovative CRF model cooperated with locality-consistent dictionary learning.Specifically,we design two locality-consistent dictionary learning strategies to capture the local consistencies in the feature/label space.We further develop a joint learning algorithm for the dictionary and the CRF model parameter.In practice,obtaining the satisfied ground truth is not only time-consuming but also labour intensive,which declines the generality of the fully supervised methods.To tackle with the issue,we also propose a novel CRF based framework for weakly supervised semantic segmentation.Enlightened by jigsaw puzzles,we first merging superpixels of an image into pieces.These pieces are then associated with proper semantic labels by CRF.Hence,a piece library is constructed,achieving universality and flexibility.In the testing phase,we compare the superpixels of a testing image with the pieces in the library and distribute them the labels that minimize the energy.Extensive experimental results on several popular databases demonstrate that our frameworks outperform or are comparable to state-of-the-art segmentation methods.Specifically,the fully supervised framework achieves better discrimination while the weakly supervised one obtains more distinct boundaries.