Class-Incremental Learning Algorithm And Its Application In Image Classification

Deep learning algorithms have made significant achievements in image classification tasks,which depend on the classification results obtained through training using all data from known classes.When new data arrives,the old data needs to be retrained.However,due to memory and privacy issues,old data is not always available.If trained with new data alone,the model forgets the old knowledge.The class-incremental learning method is designed to solve the catastrophic forgetting problem.Therefore,this paper conducts research on class-incremental learning methods in the field of image classification.The main research contents are as follows:First,to address the problem of class confusion between tasks in class-incremental learning methods that do not store old data,a class-incremental learning method for generative classifiers based on class enhancement is designed.This method enhances the generative classifiers by splicing class labels at the input of each variational auto-encoder,making the retained information of new and old classes more distinguishable,thereby reducing confusion between classes.Secondly,aiming at the problems of stability-plasticity and output bias in the class-incremental learning method with small amount of old data,a class-incremental learning framework based on double weight correction network is designed.This framework constructs stable blocks and plastic blocks on each residual layer,and use balanced subsets to adjust aggregate weights end-to-end to balance the weights of these two blocks,namely,balance stability and plasticity.At the output layer of the model,the weights of the new and old classes are corrected by the weight alignment algorithm to balance the outputs of the new and old classes.Finally,aiming at the unbalanced distribution of new and old class data in class-incremental learning methods that store a small amount of old data,a K-means based method for extracting old class exemplars is designed.Using K-means clustering to select K representative samples for each class to improve the diversity of old class data,and then using an end-to-end sample optimization framework to parameterize the selected samples according to image size,optimizing them in an end-to-end manner to better fit the distribution of the category,thereby enhancing the representativeness of old class samples and solving the problem of imbalanced distribution of new and old class data.