Incremental Learning With Label Noise For Face Recognition

Most methods of face recognition use amounts of corrected labeled samples to learn recognitionmodels with high curate. Collecting face images and labeling them will consume plenty of manpower.Further more, in real life, the impact of lighting needs classifiers learn new knowledge continually. Toachieve this acquirement, researchers advance incremental learning and semi-supervised learning: Theformer one makes use of incremental data set to improve the initial data set then mix them, the mixedsets preserve the useful part and delete the most useless data, making classifiers train continually; Thelatter one utilizes the unlabeled samples to make the accurate approach the result of labeled ones.However, when the increment set contains noise, these methods have drawbacks. The paper is suitablefor face recognition with noise, making use of corrected labeled initial set and increment set withnoise to learn. Give confidences to incremental set; delete noisy data according to confidences, thenmix the disposed incremental set with initial set to study multi classifiers which take different policiesto vote for the test set, deciding their finial classes. Main results include:(1) Multi classifiers approach based on SVM, namely Multi_SVM, is proposed. Multi_SVM trainsmulti classifiers based on SVM, discards samples judged as noise and decides test samples’ classesaccording to three vote manners. All the classifiers have tolerance limits to the noise; the limit’s heightis one of the key guidelines. The experiments prove that Multi_SVM has a better tolerance to noise.(2) Multi classifiers approach based on TSVM, namely Multi_TSVM, is proposed. Multi_TSVM’sidea is the same as Multi_SVM, discarding samples which were judged as noise, and then studying.The difference is that in Multi_TSVM, the training of classifiers is based on TSVM. The experimentstestify that Multi_TSVM improve Multi_SVM further.(3) compare Multi_SVM、Multi_TSVM and TSVM. From the experiments we get that these threeapproaches have better performance than SVM and Multi_TSVM is better than Multi_SVM.