Structural Semantic Description And Feature Selection For Image Semantic Understanding

In Recent years, image understanding has been a hot problem in computer vision.This article focuses on two respects to analysis this problem: i). Propose a high-level imagedescriptor, which based on object clique obtained by calculating the dependence degreebetween objects. This high-level descriptor contains rich semantic information, thus canhelp to improve the scene classification accuracy. ii). Propose the feature selection modelsbased on high-level semantics to select the most discriminately features for specific scene,thus achieves better recognition results in object recognition and scene classification.For i), This article proposes an image descriptor which is based on object cliques.Compare to use single object as attribute, this descriptor contains richer high-level seman-tic information, thus can help narrow the semantic gap efectively. For ii). This articlepropose two feature selection models: Spatial Path Coding (SPC) and Object Coding onthe Semantic Graph(OCSG) to induce a process of supervised structured sparse feature s-election. SPC is a regularized penalty which encodes the spatial correlations of featuresobtained by the spatial pyramid model. In SPC, each dimension of features is considered asa vertex in a Direct Acyclic Graph (DAG), and the spatial correlations among features areconsidered as the directed edges associated with the predefined weights. Thus, the processof supervised feature selection can be directly formulated to solve a minimum cost pathselection problem. Classification results and recognition results compared with the state ofthe art algorithms show the better performance of the proposed method. In OCSG, priorknowledge is first exploited by making statistics on a large number of labeled images andcalculating the dependency degree between objects. Then, a graph is built to model thesemantic correlations between objects. This semantic graph captures semantics by treatingthe objects as vertices and the objects afnities as the weights of edges. By encoding thissemantic knowledge into the semantic graph, object coding is conducted to automaticallyselect a set of object cliques that have strongly semantic correlations to represent a spe-cific scene. The experimental results show that the Object Coding on semantic graph canimprove the classification accuracy.