Classifier Design And Its Application On Object Recognition In Image Sequences

With the development of computer technology,computers have been endowed with important responsibility because it's capable to replace the human's work intelligently and efficiently.And computer vision is developed by leaps and bounds recently.As an important subject in the field of computer vision and pattern recognition,object recognition has received extensive attention.Object recognition is the process of identifying a specific or interesting object from background image based on the relevant image information.With the development of imaging platforms and technology,video sequences captured by motional or stationary platforms are becoming the common ways to acquire data.Therefore,object recognition for the sequence image data is becoming one important subject with great value.The main work of this paper is as follows:1.Based on a massive research of literatures about object recognition,this paper reviews the object recognition methods for sequence image data,and emphatically summarizes the key techniques and main ideas of object recognition algorithms based on statistical learning model.This work has laid a theoretical foundation for the future object recognition method with a higher performance.2.A object recognition algorithm under a small number of samples is proposed based on multi-classifier interaction,which solves the problem of the single classifier with a long training time consuming under the condition of a large number of samples.And it maintains a high object recognition rate.Firstly,a small number of labeled sample sets are obtained,which are trained by Bayesian,BP neural network and SVM classifier.Secondly,the primary and secondary classifiers are determined.Thirdly,the unlabeled samples are classified by the auxiliary classifier through voting method.Fourthly,the selected samples are added to the classified samples for the continuing the training and learning.The three main classifiers are selected in turn.The process above is repeated until the maximum iteration.In the end,three classifiers are integrated to obtain the final object classifier.Experiments using visible light sequence image data verify the effectiveness of the proposed method and shows that the recognition rate has been improved.3.A new object recognition algorithm based on improved AdaBoost classifier is proposed to shorten the training time and improve the recognition rate.Firstly,the haar-like feature is extracted from the samples.In each iteration,the eigenvalues are sorted in ascending order,the candidate classification set is determined and the classification position with the lowest classification error rate is found.Then,based on the classification position,a reasonable classification threshold is set up according to the distribution of the samples and the classification rules are determined.Finally,the classification rule which minimizes the classification error rate is chosen as the weak classifier of the iteration.The experimental results show that the algorithm converges and the recognition performance is improved.