An Improved Algorithm Of Active Learning Based On Multiclass Classification

For supervised learning model,enough labeled samples are the prerequisites for obtaining high precision classifiers.But in practice,unlabeled samples are relatively large in general case,and the manual annotation is very expensive.Therefore,it is necessary to control the quantity and quality of training sample sets.It is a key problem for the active learning that how to select the unlabeled samples with high classification contribution and add it to the existing training set to improve the accuracy and robustness of the classifier.Additional,most active learning research is limited to the closed sample set.It is worth studying how to apply active learning to the actual production environment and achieve high classification accuracy.In order to solve the problem of inter class equilibrium and outliers in the BvSB sample selection algorithm,a Center+reBvSB sample selection algorithm is proposed based on the combination of uncertainty and representation.Firstly,K-Means clustering is used to select the representative training set A,and then reBvSB sample selection algorithm is used to select the representative edge equalization sample set B,and finally integrate A and B and update the training set.The experimental results show that the algorithm can help the classifier to improve its classification accuracy and robustness.The Center+reBvSB sample selection algorithm is integrated into the active learning algorithm,and a new active learning algorithm based on improved BvSB is proposed.Using the misclassified sample pool generated by on-line recognition to re-train the classifier with the original sample pool,it can further enhance the recognition ability of the classifier.The experimental results show that the improved active learning algorithm has good robustness and high recognition precision.