Active Learning in Image Classification: A Comprehensive Review and Analytical

Abstract

Active Learning (AL) has emerged as an effective strategy for improving image classification performance while minimizing the cost of manual data annotation. By selectively identifying the most informative and uncertain samples for labeling, AL reduces the dependency on large labeled datasets and enhances model efficiency. This paper presents a comprehensive review of active learning techniques applied to image classification, including uncertainty-based, query-by-committee, diversity-based, and hybrid approaches. It critically evaluates their performance across various datasets and deep learning architectures. Additionally, the paper discusses key challenges such as scalability, selection bias, and integration with modern deep learning pipelines. Future research directions are also highlighted, emphasizing the need for adaptive, efficient, and robust active learning frameworks for real-world applications.