In machine learning， data quality has a far-reaching impact on the accuracy of system prediction. Due to the difficulty of obtaining information and the subjective and limited cognition of human， experts cannot accurately mark all samples. And some probability sampling methods proposed in resent years fail to avoid the problem of unreasonable and subjective sample division by human. To solve this problem， a label noise filtering method based on Dynamic Probability Sampling （DPS） was proposed， which fully considered the differences between samples of each dataset. By counting the frequency of built-in confidence distribution in each interval and analyzing the trend of information entropy of built-in confidence distribution in each interval， the reasonable threshold was determined. Fourteen datasets were selected from UCI classic datasets， and the proposed algorithm was compared with Random Forest （RF）， High Agreement Random Forest Filter （HARF）， Majority Vote Filter （MVF） and Local Probability Sampling （LPS） methods. Experimental results show that the proposed method shows high ability on both label noise recognition and classification generalization.