In the traditional AdaBoost algorithm, there are over-fitting problems caused by noise samples. In this paper, an improved AdaBoost algorithm based on noise detection is proposed, called NAdaBoost. According to the traditional AdaBoost algorithm, in the misclassified samples, noise samples vary widely in some attributes. NAdaBoost can, instead, determine the noise samples based on this, and then reuse the algorithm to classify the two types of samples, and ultimately achieve the purpose of improving the accuracy of classification. The experiment on the binary classification shows that the proposed algorithm has a higher classification accuracy compared with the traditional AdaBoost algorithm, as well as relative improvement of algorithms.