Image Feature Extraction And Recognition Based On Rich Thresholding

Computer vision has been widely used in visual surveillance, biometrics,human-machine-interfaction, industry inspection. The core of such computer visionsystem is object detection and recognition while the main stages of object detectionand recognition are feature extraction and classifier design. Feature extraction ischanllenging for visual images due to large vartions in illumination, occlusion, pose,and background. Therefore, feature extraction is investigated in this thesis. It isexpected that the features are roblust, discriminative, and compact. Feature extractionis followed by classifier design. Two main factors of classifier are generalizationability and efficiency. Generalization ability is the ability of a classifier that cancorrectly detect and recognize objects while the efficiency of a classifier determineswhether the computer vision sysmtem can work in real time. So this thesis also placesemphasis on efficient classifier design.In the aspect of feature description, this paper proposes two novel imagedescriptors base on LBP, which are double local binary pattern (DLBP) and localquaternize pattern (LQP). One of the drawbacks of the traditional LBP is that muchspatial information is discarded. The proposed two methods attempt to overcome thedrawback by taking much more spatial information into account, DLBP takes anotherLBP processing for the LBP image, then combines histograms of two LBPs together,and LQP takes not only the LBP between the center pixel and the smallerneighborhood, but also the LBP between the smaller neighborhood and the biggerneighborhood, then combines them.To extract compact low-dimension visual features in image search reranking task,a new method named RNPE-Reranking is proposed. Firstly, the proposed methodintroduces the ranking relevance information to Neighbor Preserving Embedding(NPE) algorithm, and presents a new dimensionality reduction algorithm calledRelevance Neighbor Preserving Embedding (RNPE). And then, a typical rankingalgorithm, RankingSVM, is applied to PNPE to obtain a ranking model. Finally, theinitial images are reordered with the ranking model.Moreover, it is found that exisisting cascade AdaBoost algorithm costsunnecessary computation on classifying positive sub-windows. To deal with this problem, this thesis proposes a novel scheme that utilizes two thresholds for efficientclassification. It is the threshold designed for positive response that makes it possibleto classify positive sub-windows at earlier stage.A common characteristic of the above methods is that additional but importantfactor is introduced in the methods. In the proposed double local binary pattern(DLBP) and local quaternize pattern (LQP), not only information of second order butalso four discrete values are used. In the proposed RNPE-Reranking algorithm, thevalues of different ranking scores are incorporated. In the proposed cascade AdaBoostalgorithm, a important threshold is used to efficiently accept positive sub-windows.