An improved fusion method based on Adaboost algorithm for semantic concept extraction

In this paper, based on probability distribution of weak classifier output, an improved Adaboost-based multi-classifiers fusion algorithm is proposed for semantic concept extraction. We present a novel method to compute the error rate and the weight of each classifier. We believe that the error rate of an example should be related to its rank in a weak classifier output. First, the probability distribution of the SVM output is estimated. SVM is regarded as the weak classifier in our system. Then, based on the negative and positive examples probability distributions, we can calculate the error rates of positive and negative example respectively. We define the error rate of a positive example as the proportion of negative examples whose scores are bigger than this positive example in an SVM output. Finally, we integrate the error rate into the Adaboost algorithm and add some modification to further improve our performance. We call the proposed fusion method D-Adaboost since the distribution-based error rate computing algorithm is integrated. Experimental results on TRECVID-2007 dataset show the effectiveness of the proposed D-Adaboost.