Image Parsing With Spatial Context

Scene understanding is the main difficulty and hotspot in the field of computer research.Its fundamental goal is to make the computer as the human brain understand the natural scene.The development of computer vision and artificial intelligence has promoted the progress of image understanding.After decades of research,image understanding technology has developed by leaps and bounds,but there are still very challenging problems,such as the large number of objects in the scene.In natural scenes the ocean and sky occupy most of the scenes,while people and cars are very small parts,which often express the key semantic information,which makes the scene semantic understanding is greatly hindered.For scene understanding,it is very important to make full use of global context information.Many existing models or methods are only limited to a small range of context information modeling,only learning local features.Aiming at the limitation of local context,this paper proposes an algorithm of spatial parsing with spatial context(PSC).The spatial relation dictionary is constructed by extracting the position feature vector to describe the positional relation and the co-occurrence of the objects.The spatial context information is added into the Markov random field to supplement the global spatial context information.In the context of the expression of information,the image of different scales of the division,extraction of features in the superpixel.PSC has been experimented on MSRC and SIFT Flow datasets.Compared with the most advanced algorithms,PSC has an absolute superiority in the accuracy of object classification,and the accuracy of pixel labeling is higher than that of most other algorithms.In summary,this paper proposes a spatial context-based algorithm for scene understanding,which captures the global information from the perspective of semantic and spatial context information.It uses high-level semantic features to describe scene content,which is of great significance to the research of scene understanding.Experimental results on challenging data sets show that the proposed algorithm outperforms the latest algorithms in terms of accuracy and robustness.