Unlabeled data selection for active learning in image classification.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Active Learning (AL) addresses the challenge of extensive data labeling in data-intensive fields.
• AL aims to identify valuable unlabeled data samples for annotation using acquisition functions.
• Image classification in computer vision requires substantial training datasets.

Purpose of the Study:

• Introduce innovative AL selection methods for image classification.
• Identify informative images from unlabeled datasets.
• Minimize the amount of training data required for effective model training.

Main Methods:

• Developed Similarity-based Selection, Prediction Probability-based Selection, and Competence-based Active Learning methods.
• Evaluated proposed methods on Cifar10 and Cifar100 datasets.
• Compared novel methods against random and conventional selection techniques.

Main Results:

• Proposed AL methods demonstrated superior performance compared to random selection.
• Novel techniques outperformed conventional data selection approaches.
• Effectiveness in enhancing the AL process for image classification was proven.

Conclusions:

• The novel selection methods are effective for Active Learning in image classification.
• These methods reduce the need for extensive data annotation.
• Significant improvements in the AL process for computer vision tasks were achieved.