Studies On Efficient Online Incremental Learning For Discriminant Analysis

Many real-world applications of high-dimensional large-scale images involving visual tasks,such as target recognition,tracking,image segmentation and labeling,all require to get distinguishable feature from continuously arriving data streams.It is very common to have a small amount of data available in a short period of time.With the migration of time,there are more data available.Therefore,it is necessary to add these data to the trained model in an incremental way instead of retraining.These newly arrived data include new samples with or without labels.So it is challenging to update model step by step in an efficient and stable manner,especially for high-dimensional or large-scale visual block data streams.Linear discriminant analysis(LDA)has good performance in the exploration of data features,and it can cut down computation load in a degree by reducing the dimension size of data to a large extent.However,most existing methods need large computation time and storage space requirements to get the optimal projection matrix of the LDA method.In addition,most of them only consider the adding of inserting new data of existing and new class with known labels,while few consider labelled an unlabeled mixture data.In this paper,three kinds of fast LDA learning methods and their incremental versions are proposed in order to solve the problems of incomplete incremental situations,large amount of calculation and large storage space.These three algorithms are based on LDA,and they gradually consider more incremental situations and improve the efficiency of the algorithm.The first method proposes an incremental learning algorithm based on null-space LDA with twice QR decompositions for learning new samples and new classes in the zero-shot learning model based on attributes.The second method proposes a fast factorization-free algorithm by taking the class centers as the model input,and using the product of the class centers instead of the QR decomposition.For inserting the new samples,new classes,mixed data with labels into model,the corresponding incremental learning methods are proposed,which improve the computational efficiency.The theoretical analysis and experimental verification results show that this algorithm is effective in real databases.The third method uses kernel skill into the second method,and proposes a method for novel class detection with a fixed threshold according to its unique characteristics in the case of unlabeled mixture data streams.In recognition rate and computational time,it shows its superiority.