Research On Image Incremental Learning Based On Deep Learning

With the development of computer vision and artificial intelligence,people need more flexible strategies to deal with the large-scale and dynamic classification in the real world.At least,when a training data of new categories is introduced,the artificial neural network should be able to incrementally learn the features of the new category rather than relearn all the data.The incremental model can save a lot of time cost for the training of artificial neural network,and guarantee the high performance of the model.However,incremental learning meets the above requirements.Through split the big data into stages,it not only saves a lot of training time cost,but also meets the high demand for accuracy.However,the problem of catastrophic forgetting exists in incremental learning,and the incremental learning algorithms are generally limited by the over-dependence on system memory and the huge network architecture.Under this background,this paper starts to study the incremental learning algorithm based on the dynamic correction vector with deep learning models.The main contents of the paper are as follows:(1)Incremental model research based on deep neural network.The backbone network of the model adopts Alex Net and Res Net,which are recognized as effective deep neural networks,to ensure a strong ability to identify each image dataset.The incremental technology based on deep neural network mainly includes the following steps: first,the representative memory is proposed as a method of storing the memory of old classes,and training the new class in the subsequent incremental stage together to ensure the recognition ability for the old and new class.Second,the memory and computing resources of the system should be bounded.To specify the memory upper bound of the system,a hyper-parameter K is proposed,called the memory budget of the classifier.Finally,in each incremental process,the total loss is calculated by the classification and distillation loss using the new and old data stored in the representative memory.(2)Research on incremental learning algorithm based on the dynamic correction vector.The techniques of representational memory and knowledge distillation create new problems called bias problems and overfitting problems.For the existing categories,knowledge distillation will cause obvious noise in the teacher model and mislead the student model.The unbalanced training dataset generated by representative memory may lead to the inaccurate feature representation in the model.The correction vector represents the statistics of the probabilistic label and the ground truth label in the training process.The classification results are usually biased towards the class with more data participating in the training.Therefore,the correction vector intuitively reflects the degree to which the model is biased towards each class.Since the data changes dynamically at each incremental stage,this paper proposed a dynamic correction vector to be applicable to the incremental learning.(3)Evaluate and analyze the proposed incremental model.In order to evaluate the incremental model of Alex Net and Res Net in image and signal datasets,this paper compared proposed incremental algorithm with the three best incremental algorithms in the academic field,and evaluate the performance of every algorithm in terms of final accuracy,average accuracy,confusion matrix and incremental training time cost.It is proved that the proposed algorithm has generalization in each dataset,and keeps the highest accuracy and the shortest training time cost.(4)Incremental learning applications based on fixed features.Because signal data is often invisible,it often takes professionals to distinguish it,resulting in significant labor costs.Therefore,in the context of signal data,incremental recognition of signal data is very necessary.Through the method of fixed features,the characteristic library of signal data is established,which can save the time cost of professionals and improve the work efficiency.