Research On Strategies Of Active Learning And Its Application To Image Classification

Image classification is a cross area of computer vision,pattern recognition and machine learning.It aims to help machine extract discriminative feature from images or image sequences and classify them.With the development of big data technology and computing power,deep learning has made remarkable achievements in many fields,which has been widely applied in medical image processing,intelligent trans-portation,e-commerce platform,face recognition and so on.The powerful ability of deep learning is highly dependent on a large number of labeled samples.However,most of the real-world application scenarios suffer from the shortage of annotated samples,which requires considerable human efforts.Recently,as an emerging tech-nology,active learning has attracted great attention,which can observably reduce the cost of annotation using some heuristic strategies.Therefore,research on strate-gies of active learning and its application to image classification is a challenging and meaningful work.The main works of this thesis are as follows:Firstly,we summarize active learning methods and image classification tech-nologies based on lots of research,including 1)traditional machine learning and deep learning based image classification technologies.2)the basic framework of ac-tive learning,some relevant strategies and its extension methods.3)the evaluation index of experiments,databases for evaluation and experimental environment.Secondly,we propose a framework for taking full advantage of limited resource that interactively integrated semi supervised learning and batch mode active learn-ing,named NRMSL-BMAL.1)We propose a memorized self-learning(MSL)algo-rithm to reduce noisy labels owing to the memorized information of historical pre-diction.Furthermore,to reduce the impact of the remainder noise labels,we apply a noisy label self adjusting method in an effective way to MSL,named NRMSL.2)To improve the efficiency of BMAL,we combined a convolutional autoencoder based cluster algorithm and uncertainty strategy.3)We performe a thorough experimental evaluation in image classification tasks and show empirically that NRMSL-BMAL could observably reduce annotation cost.Thirdly,we systematically analyze the series of generative adversarial networks and adversarial autoencoder.Considering the quality of generated images and its application to active learning,we propose a generative adversarial network based two stage active learning method.Experimental results in image classification datasets show that our proposed method could effectively improve the quality of generated images and reduce the cost of annotation and time.Finally,by considering the real application scenarios of active learning,we de-sign and implement an active learning system for image classification tasks.Active learning system has been simulated to verify its availability.