Image Classification Based On Active Learning

With the development of multimedia information technology, the number of digitalimages has been growing explosively; automatic content-based image classification hasbecome a hot research area which mainly includes two aspects as feature extraction andtraining of classifier. Obtaining the accurate classification model often requires a lot oflabeled samples which are costly to get while the promise of active learning is just this:when some unlabeled examples to be labeled are selected properly, the number oflabeled samples for some problems decreases drastically so the cost of manuallylabeling can be reduced. This thesis first summarizes the techniques about image featureextraction and classification algorithm, and then studies the combination of activelearning with image classification. The main work of this thesis is as follows:1) The first part summarizes the commonly used features in image classificationtasks, including low-level features, mid-level features and high-level semantic featuresand giving theory details of local descriptors Scale Invariant Feature Transform. Thispart mains focus on analysis of extraction process of the representatives of the mid-levelfeatures: Bag of Words and sparse coding.2) The second part summarizes the common image classification algorithms,including Bayesian decision theory, support vector machines, Logistic regression andrandom forests, etc. This part presents some images classification experiment’s resultscombining with the mid-level expression. To satisfy engineering need, onedemonstrated software was written to present the overall process of image classification.3) The third part summarizes the representative algorithms of active learning,including the theoretical basis and scope of application.4) Finally and the most important part, to enhance the efficiency of active learningprocess, this part introduces incremental support vector machine in the classifierretraining process for active learning. To reduce the redundant information among batchselected samples, a multi-pool (Multi-pool) method has been proposed based on affinitypropagation (AP) clustering technique. The experiment results have verified themethod’s effectiveness on three public image datasets algorithm than traditional activelearning strategies.