Study On Image Data Streams Classification Based On Online Semi-supervised Learning

In the digital age,a large amount of data are being generated all the time.Existing supervised methods are difficult to process these data.For example,in handwritten digit recognition tasks.The methods for image stream classification have some shortcomings,that is,they either require a large amount of labelled training data,or it is difficult to adapt to the new incoming data.Therefor,it is particularly important to design an efficient algorithm that can handle concept drift and label scarcity for streaming data.This thesis focuses on the classification of image data streams based on online semi-supervised learning.The main contents are as follows:1)For the high efficiency requirements of image classification,LLDRB algorithm is proposed.The proposed algorithm models with limited labelled data.The incoming data will be predicted by the model and used to upadatae it.The whole process is transparent and interpretable and the model can be trained online on a sample-by-sample or chunk-by-chunk basis.Numerical examples based on large-scale benchmark image sets demonstrate the strong performance of the proposed classifier as well as its distinctive features compared with the “state-of-the-art” approaches.2)A new pruning method is proposed to solve the problem of of limited memory resources.When a fuzzy rule contains too many data clouds,the edge data clouds are removed by our proposed method by calculating the global density of this rule.The pruning method can ensure that the number of data clouds are always in a certain range.In this way,the algorithm can be used for life-long learning with limited memory.3)In order to make the classification result more accurate,the concept of "data cloud gravity" is proposed.For an uncertain data,the classifier will calculate the gravity between the neighboring data clouds and the uncertain data with their qualities and distances.By comparing the gravity of each rule,we will get a finer result.The experimental results also prove the validity of "data cloud gravity".Numerical examples based on large-scale benchmark image sets demonstrate the strong performance of the proposed SSDRB classifier as well as its distinctive features compared with the “state-of-the-art” approaches.