Research Into Multi-category Recognition Algorithm Based On The Tree-structure Model

Many Research areas are related to multi-category pattern recognition, which has a wide range of applications. Meanwhile, it is rather challenging. We intend to propose a popular algorithm on multi-category pattern recognition. Moving object recognition in video is an important issue in the field of computer vision; cell mitotic phase identification is of great significance in the field of medicine and biology. In order to solve these two typical problems related to multi-category pattern recognition, based on the summary and analysis of the relevant research, we improve the decision tree algorithm and learn classifiers through the SVM algorithm and the AdaBoost algorithm. We propose a classification algorithm that we build a tree-structure model based on feature(classifier) selection strategy.We integrate SVM algorithm and AdaBoost algorithm into the decision tree algorithm, and learn classifiers through this two algorithms in order to improve single classifier's performance. For moving object recognition, we divide all features into two categories: numeric features and descriptive features. For low-level numeric features, we learn classifiers through the AdaBoost algorithm. For high-level descriptive features, we learn classifiers through the SVM algorithm. For cell phase identification, we extract features of different phases'cell and learn classifier, and we obtain collection of classifiers.The tree model's structure is determined by the segmentation of samples on tree node. We build the tree-structure model based on the feature(classifier) selection strategy. Prior knowledge are introduced; in the current node the importance of all classifiers are determined by testing the classifier's performance and distribution of all samples. Feature selection is based on the fact that the best classifier is chosen and placed to the current node by sorting all classifiers'importance. In turn for current node, we segment the samples through feature selection to extend tree structure gradually. Each tree node's classifier is approximately considered as a second-class classifier. In this thesis, multi-category classification problem is transformed into a series of second-class classification problems through the tree-structure model's multi-storey decomposition strategy.The tree-structure model is build based on feature selection. Compared to the manual-designed model it has great flexibility and rationality. Based on the tree-structure model, we carry out experiments for moving object recognition and cell phase identification. Compared with the similar framework quantitatively, our approach shows better performance. At the same time, due to its universality, our algorithm can help to solve other multi-category pattern recognition problems.