Incremental learning using error and sensitivity analysis of MCS for Image classification

As the Internet refreshes every day, a large scale of images are generated online which present a challenge to image classification problems. Firstly, the classifier once trained by the old training set is not able to describe all the characteristics of a class when new samples appear. Secondly, to train a classifier using all the upcoming samples can take a long time so that the speed of updating the classifier is much slower than the speed of new data generation. Thirdly, the newly generated images may be duplicate or similar to current training samples with minor variance, hence training by these minor informative images will waste lots of time and resources, Samples being continuously misclassified by the updated classifiers should be laid with more weight in future update process than other easily classified samples. In this paper, we propose an Incremental learning method using Error and Sensitivity Analysis (IESA) of Multiple Classifier System (MCS) for upcoming images. Radial Basis Function Neural Network (RBFNN) is used to classify upcoming images firstly and misclassified images with large sensitivity are selected for the following updating process. Experimental results on a large scale image dataset convince the efficiency of the IESA strategy.