Research On Class Incremental Image Recognition Method Based On Feature Replay

With the rapid development of deep learning technology,image recognition algorithms based on deep neural networks have achieved excellent performance.However,most of these methods are based on the batch learning paradigm,which assumes that all data is obtained at once before training.In dynamic open scenarios,data is often characterized by the arrival of streams,which is time-varying and infinite.Therefore,incremental learning came into being,the main idea of which is to enable models to continue to learn like advanced creatures such as humans,deal with new tasks that continue to come,and keep the old tasks unforgotten,so as to solve the problem of high dependence of batch learning paradigm on complete data.Most of the existing incremental learning methods review old knowledge by saving old class samples or generating pseudo samples,which is difficult to apply in scenarios such as limited storage and data privacy.The feature playback-based method effectively reduces the large storage and computational overhead required by the sample playback method,but it still has shortcomings in the process of long sequence increments.This paper mainly studies the incremental learning problem in the visual task of image recognition,especially the challenging class incremental learning problem.In this paper,the following research is carried out on the class incremental learning method based on feature replay:(1)In this paper,the problems of prototype drift and cross-task normalization differences in the prototypes replay method in the long sequence increment process in the feature playback method are analyzed,and a method based on multi-level distillation and continual normalization is proposed.It is improved from the aspects of distillation strategy and network backbone,and the multi-stage distillation framework is used to constrain the knowledge transfer between new and old models and reduce prototype drift.In addition,a continual normalization layer is introduced into the network backbone to overcome the cross-task normalization differences existing in the traditional batch normalization layer to adapt to non-stationary incremental data.Experiments on the CIFAR100 and Tiny Image Net datasets show that compared with the benchmark method,this method effectively improves the average incremental accuracy,and significantly reduces the average forgetting rate.(2)Prototype replay approaches typically utilize knowledge distillation strategies to complete knowledge transfer between old and new models,which require the construction of at least two models(previous stage model and current stage model).In addition,constraints based on distillation loss often lengthen the model training cycle.Based on the method of feature generation,this paper further explores the realization of long sequence incremental learning under a single model,and proposes an embedding expansion module to assist in learning smooth embedding space,using graph structure and embedding interpolation technology to obtain richer feature embeddings,and then training the joint classifier on the extended feature embedding,so as to enhance the robustness of features to noise and improve the accuracy of model classification.Experiments on the CIFAR100 and Tiny Image Net datasets show that the method effectively improves the average incremental accuracy,and significantly reduces the training time and the amount of model training parameters compared with the distillation-based prototype replay method.