Boosting Approach And Application Research On Image Understanding

Image Understanding is the hotspot in computer research area. Main research contents are object recognition in scene and scene description and understanding. Classification and detection are key problems of object recognition in scene, which produce discriminative decision through training and learning to obtain recognition results. Whether discriminative model is good or not directly influences recognition results. Boosting can combine weak classifiers that perform just slightly better than random guess into an accurate strong classifier, which can solve more general object recognition in Image Understanding.This dissertation includes the following contents:(1) The dissertation discusses the classification mechanism of Boosting, and summarizes the two classification models: AdaBoost and Boost-by-Majority. It also classifies and compares spreading models from four aspects: weights of training samples and weak classifiers, loss function and characteristics, forming a comparatively complete system of Boosting.(2) Statistical nature of Boosting is explained, mainly from two aspects: exponential criterion and log-likelihood criterion. Then Gentle AdaBoost is compared with Real AdaBoost through data classification experiment, which verifies advantage of Gentle AdaBoost so as to lay a theoretical foundation for model selection.(3) Traditional image database and LabelMe database with visual knowledge are analyzed. Personal image database which contains specific object class and is suitable for object recognition in scene is created by adding annotations to them.(4) Meaning and existing methods of object recognition in scene are summarized. Then object recognition in scene based on Gentle AdaBoost is realized, which judges whether images have specific object class or not and determines the location of object. Finally, evaluation of recognition results via Recall Precision Curve is given.