Adaboost Algorithm And Its Application Research In Object Recognition

With the rapid development of computer hardware technology and imageprocessing technology，visual target recognition and tracking technology is widelyused in the military field，aerospace，scientific exploration，astronomical observationand civilian areas. Target recognition and tracking technology has become a hot topicof research in the automatic control field， computer vision field and Patternrecognition field. It presents mainly on target recognition and tracking technologywith practical research background.The learning method of discriminated model mainly including: Ensemblelearning methods, statistical learning methods and the organization learning method,etc. AdaBoost algorithm is one of Ensemble learning methods, it has been applied to awide range of object recognition. The basic theory of AdaBoost algorithm is to selectseveral weak learners which has just slightly better accurate than random predictionand combines them into a strong learner. This paper focus on: unbalance datasetprocessing and misclassification samples process in Gentle AdaBoost.The dissertation includes the following contents:First of all, the dissertation introduces the classification mechanism of Boostingand AdaBoost. The Gentle AdaBoost is compared with Real AdaBoost through UCIdata classification experiment, which verifies stability of Gentle AdaBoost.In the second part, Traditional method always use over-sampling way toaccomplish the implementation of minority samples for the purpose of achieving thebalance of data set, and obtain the classifier through training the samples in theprocess of dealing with the classified issues of unbalanced data set. But this methodwill incorporate the singular sample which is hard to classified, and lead to theunsatisfied classification performance of the classifier. Therefore, this paper proposesa improved Gentle AdaBoost algorithm specified for the classified issues ofunbalanced data set. Firstly, considering the feature that misclassification samples isassigned with a large weight when the classifier based on Gentle AdaBoost algorithmin training, we can decide the weight threshold for the copy samples; then, copy anumber of minority samples in the threshold range, and use the aforesaid data set totrain the classifier and obtain related weak classifier. Repeat the former procedures to balance the data set so that the strong classifier can be also obtained. The wholeprocess has the capability of avoiding the issue of incorporating singular samples inthe process of data over-sampling. The experiment demonstrates validity of ouralgorithm.Finally, we proposed an Over-sampling Algorithm Based on Gentle AdaBoost inUnbalanced Data Set. In order to solve the problem of classification performancedecrease caused by samples which difficulty to classifier, in the Over-samplingprocess. Algorithm, firstly, calculates the number of minority class and majoritysamples, and then we copy the minority class samples to make two types data samplesbalance. The paper use Gentle AdaBoost Algorithm to train classifier, and set athreshold as standard to judge which is reliable samples, then copy the samples tomake two types data samples balance, and improve its classification performance...