Abstract:Active learning (AL) aims to extract the most informative samples for labeling, and subsequently, classifier training, and thus it is deemed as a useful technique for image classification tasks. In data-scarece regions, especially those with difficult human accessibility, the usefulness of AL tends to increase since in such areas it is meaningful to achieve sufficiently high classification accuracy with a limited number of training samples. The success of an AL scheme is largely dependent on its capability of quantifying the informativeness of the unlabeled candidate samples. Although a number of studies exist for AL optimization, there is still room for improvement when it comes to designing an accurate and unbiased AL criterion. This study develops a new AL technique by using binary support vector machines (SVM) and then combines the new uncertainty metric with a clustering-based representativity diversity model and a cuckoo search method. The contributions are mainly two-fold. First, to explain the advantage of the proposed uncertainty criterion, the traditional SVM-based approach is described and its defects are demonstrated. The proposed uncertainty metric is then introduced by illustrating how the shortcomings can be avoided. Second, the new uncertainty model is combined with a clustering-based representativity diversity, and a cuckoo search algorithm is used to further diversify the selected samples. To validate the new AL approach, a multi-spectral image of high spatial resolution captured over a coal mining area in northern China is adopted. The experimental results indicate that our approach can achieve superior classification accuracy when it is compared to several competitive methods. When the initial configurations are the same, our scheme can obtain an overall accuracy of 93.95%, which is higher than those of other competitors.

Combination Of Competitive Em Algorithm And Support Vector Machine For Active Learning

Multi-class AdaBoost ELM

Uncertainty-Based Active Learning Via Sparse Modeling for Image Classification

Dual-Classifier Collaborative Method Based on Semi-Supervised Active Learning

Uncertainty Sampling Based Active Learning with Diversity Constraint by Sparse Selection.

Semisupervised SVM Batch Mode Active Learning with Applications to Image Retrieval

A Two-Stage Active Learning Method for Image Classification

Semi-supervised SVM Batch Mode Active Learning for Image Retrieval

Combining Clustering Coefficient-Based Active Learning and Semi-Supervised Learning on Networked Data

Hierarchical exploration based active learning with support vector machine

Combining Active Learning and Semi-Supervised Learning Based on Extreme Learning Machine for Multi-class Image Classification

Classifier Combination Based on Active Learning

Asymmetric semi-supervised boosting for SVM active learning in CBIR

Active Learning with Adaptive Heterogeneous Ensembles

Employing Feature Mixture for Active Learning of Object Detection

AL-ELM: One Uncertainty-Based Active Learning Algorithm Using Extreme Learning Machine

Active Learning with Spatial Distribution Based Semi-Supervised Extreme Learning Machine for Multiclass Classification

Active learning with adaptive regularization

Uncertainty-aware Active Learning for Optimal Bayesian Classifier

Active learning based on SVM and representativity in a coal mining environment

Efficient Semi-supervised Active Relevance Feedback Scheme for Image Retrieval