Semantic Image Segmentation Based On Deep Structured Learning

Semantic image segmentation is an important branch of computer vision and the core technology for image understanding related applications.At present,semantic image segmentation technology has made great progress,and a variety of important segmentation methods have been proposed.Among them,the methods based on deep learning and the methods based on conditional random field have attracted extensive attention.The main reasons why CRF can play such an important role include:(1)combining rich learning characteristics and manual features;(2)having relatively flexible potential energy functions(such as the Potts function)to smooth segmentation;(3)being able to explicitly model the label consistency of local regions of the image(such as "water and boat" is significantly more consistent than "road and boat").When the local features are weak,the semantic segmentation results based on CRF are significantly better than the semantic segmentation without CRF.Deep learning of the CRF graph structure is a challenging issue.First,the solution is too large to consider all possible graph structures.Secondly,because the graph structure is heterogeneous,the isomorphic structure learning method is not applicable.Again,because each particular graph structure corresponds to only a single image,the training data is lacking.This paper mainly studies semantic image segmentation based on deep structured learning from three aspects: CRF graph structure learning based on siamese network,CRF graph structure learning based on RCF network and deep structured semantic image segmentation.The characteristics and innovations of this paper mainly include the following three aspects:1.Aiming at solving the problems of excessive solution space,heterogeneous graph structure and lack of training data in the structure of CRF graph,a CRF graph structure learning method based on siamese network is proposed.2.Aiming at the problem of poor accuracy of constructing CRF graphs based on siamese networks,a deep learning method based on RCF network is proposed.Compared with the learning of CRF graph structure based on twin network,this method achieves a significant improvement in graph structure accuracy.3.After constructing the conditional random field model for semantic image segmentation based on the results of deep CRF graph structure learning,this paper proposes a gradient descent algorithm based on pseudo log-likelihood estimation strategy,which can effectively learn CRF parameters.The experimental results show that the CRF graph structure has a great influence on the semantic image segmentation performance.The CRF model constructed based on the CRF graph structure learning method proposed in this paper can obtain a significant improvement in segmentation accuracy compared with the traditional tree structure and the adjacent structure CRF model.The proposed algorithm is validated on the public and important semantic image segmentation datasets.The results show that the proposed semantic image segmentation algorithm based on deep structured learning can achieve or exceed the segmentation performance of the frontier methods.