Crowdsourcing systems provide a cost effectiveand convenient way to collect labels, but they often fail toguarantee the quality of the labels. This paper proposes a novelframework that introduces noise correction techniques to furtherimprove the quality of integrated labels that are inferred fromthe multiple noisy labels of objects. In the proposed generalframework, information about the qualities of labelers estimatedby a front-end ground truth inference algorithm is utilizedto supervise subsequent label noise filtering and correction.The framework uses a novel algorithm termed adaptive votingnoise correction (AVNC) to precisely identify and correct thepotential noisy labels. After filtering out the instances withnoisy labels, the remaining cleansed data set is used to createmultiple weak classifiers, based on which a powerful ensembleclassifier is induced to correct these noises. Experimental resultson eight simulated data sets with different kinds of featuresand two real-world crowdsourcing data sets in different domainsconsistently show that: 1) the proposed framework can improvelabel quality regardless of inference algorithms, especially underthe circumstance that each instance has a few repeated labels and2) since the proposed AVNC algorithm considers both the numberof and the probability of potential label noises, it outperformsthe state-of-the-art noise correction algorithms.