Object Detection And Recognition Based On Context

In computer vision, the most common approaches to detect and recognize object are based on local features. However, local features based methods suffer from uncertainties or ambiguities, and search complexity. Context information brings two advantages:eliminates uncertainties or ambiguities and facilitates a more efficient search. Accordingly, context is used to resolve the problems of local features based methods. This dissertation mainly focuses on object detection and recognition methods, including context features in different scales and classification algorithms.In the research of geometric context, a novel method is proposed, which extracts the ge-ometric relationship of keypoints using a set of biologically inspired templates and describes the stability of keypoints using lifetime. The proposed approach captures both photometric and geometric information, and successfully achieves an effective trade-off between gener-alization ability and discrimination ability.In the research of neighbor context, considering the importance of parts detection in the part-based representation for object detection, we present a novel part-based model which searches parts under the constraints of their neighbor information more accurately, which is described as Block local binary patter. This framework improves the speed and detection rates of object recognition.In the research of global context, we analyze the performance of global context of slid-ing windows in object detection. Then we propose a top-down object detection framework, which combines global contextual features, local contextual features and local appearance features in a coarse-to-fine cascade. The three features mentioned above play different roles in the process of object detection, and the representation with rich information makes our method robust and effective in both speed and accuracy.In research of scene context, two improves are presented. The first one is about tem-plate matching. By surveying the psychophysical and neurophysiological researches, we obtain a rule:consistency of spatial layout of scenes. Based on the rule, we present a novel approach for robust template matching, which measures the similarity of a template and a sub-image by putting both into scene context. The second one is about scene classification tasks. After the comparison of two scene context based models, we present a more robust scene classification method which integrates both pixel domain and modulation domain in-formation, and coarse spatial layout information together. The method is very effective for scene classification tasks.In the research of classification algorithms, a high-effective AdaBoost algorithm is proposed to resolve two key problems in AdaBoost, selection of weak learners and opti-mization of the strong learner, which combines a distance related discriminative criterion and optimization method using kernel-based perceptron. The two improvements ensure the efficiency of our proposed algorithm.At last, we summarize the presented work. According to the imperfect aspects, we analyze and discuss the future work.