Research On Image Annotation Based Active Learning

In recent years, with rapid development of multimedia technologies and networks, digitalimages are increasing in an explosive way. The majority of digital image are unknown data,and annotating the unlabeled samples requires a high workload and is time-consuming.Therefore, much more attention has been paid to find a good method that significantlyminimums labeling efforts. As a promising approach, active learning is often adopted toreduce the labeling efforts in classification problems. The active learner has the ability toautomatically select the most informative samples for human labeling, and then the labeledsamples are added to training datasets for building up the learner. In such iterative process, theaccuracy of the learning algorithm is improved significantly with less training samples,avoiding acceptance of the redundant informance.This paper does some research on sampling strategy and model selection. The mainworks are as follows:In uncertainty sampling, it is easy to select outliers, especially samples distributedunevenly. To tackle this issue, this paper takes into account the prior distribution of samples,and proposes an active learning algorithm based density-entropy, namely the combination ofthe uncertainty and representativeness. Neighbor entropy and neighbor density are adopted tomeasure the sample’s uncertainty and representativeness, respectively. The experimentalresults from UCI benchmark demonstrate that the proposed sampling algorithm performsbetter than random sampling in most case.In active learning support vector machine, it is a common practice to use default valuesfor model parameter, leading to unsatisfying learning performance. This paper proposes anactive learning SVM algorithm with regularization path that can fit the entire solution path ofSVM for every value of model parameters. This algorithm traces the entire solution path ofthe current classifier and find a series of candidate model parameters, then uses unlabeledsamples to select the best model parameter. The experimental results from UCI andNUS-WIDE data sets show that the algorithm proposed in this paper outperforms the defaultparameter method.