Rarity-Aware Stratified Active Learning for Class-Imbalanced Industrial Object Detection

Object detection systems deployed in industrial environments are often constrained by limited annotation budgets, severe class imbalance, and heterogeneous visual conditions. Active learning (AL) aims to reduce labeling costs by selecting informative samples; however, existing strategies struggle to simultaneously ensure robust performance, rare-class coverage, and stability under realistic industrial constraints. In this work, we propose a rarity-aware, stratified AL framework for industrial object detection that explicitly aligns sample selection with class imbalance and annotation efficiency. The method relies on a composite image-level score that jointly captures model uncertainty, informativeness, and complementary diversity cues, while adaptively emphasizing rare classes. Crucially, a stratified querying mechanism is introduced to explicitly regulate class-wise sample allocation during selection, playing a key role in improving performance stability and rare-class coverage under severe imbalance, without sacrificing global informativeness. The proposed approach operates purely at the data-selection level, making it detector-agnostic and directly applicable to modern object detection pipelines. Experiments conducted on two real-world industrial datasets involving lobster and snow crab parts, using YOLOv10 and YOLOv12, demonstrate improved training stability and annotation efficiency across balanced, imbalanced, and noisy settings over multiple active learning cycles up to 15% labeled data. Complementary comparisons with fully supervised training further show that using only 45–65% of the labeled data is sufficient to retain more than 97% of full-supervision mAP@50 and over 90% of mAP@50:95.