Study On Scene And Structure Based On Object Decection And Recognition

Scene context and structure based object detection and recognition technic is aiming at localizing the specialized object in the large image database efficiently. This paper studied the model of object existence to in the large image database and the matching algorithm while object is considered exist. Compared with the traditional approach, this paper is focus on the modeling on the relationship between the scene context, namely the relationship between feature of the scene and the existence of the object. After obtained the images those who have a high probability of containing the special object, this paper proposed a matching algorithm to get the object so as to promote the efficiency of object detection and recognition in a large image database.The studies before mainly focus on the objects'intrinsic features while the relationship between the scene context and the object is ignored. However, studies have proved that observers are capable of obtaining meaningful information from the scene with 200 msec. In this period of time, no detail is obtained by the observer. Previous tests showed that the vision system of human being mainly rely on the information of the scene. Future studies also showed that before any decision is made, the scene is categorized. Hence, the scene context plays a very important role in the object detection and recognition. By descripting of the scene, we modeled the relationship between the scene context and the object existence. However the dimensionality of the scene is very high which is very hard to model. According to the theory of the human vision system, there are five hierarchies. The information reduces dimensionalities when passing by each hierarchy. So we use some technic to reduce the dimensionality before modeling. Finally, in this paper, we use Gaussian mixture model to model the feature of the scene. Furthermore, in this paper, we discussed the relationship between position and scale of the object and the context of the scene. Images containing the certain object could be found by the object existence model, in these images, we proposed an object matching algorithm to accurately find the certain object. In this paper, we start from the graph theory, consider the feature similarity and the structure similarity meanwhile and solve the matching problem as a linear programming problem. We construct the feature which contains the information of the neighbors. Although it is not affine invariant, the approach we proposed not only is robust to 3D rotation and drape, flip and some other transformations.