Choolha Hwang  Dongha Shim

10.5626/JOK.2025.52.11.970

gemstone identification computer vision domain bias domain generalization domain adaptation image enhancement

Recent applications of machine learning and computer vision in gemstone identification increasingly utilize internet images as training data. However, commercial enhancements such as color correction, contour sharpening, and shape distortion-create visual discrepancies that result in domain bias, significantly degrading model performance in real-world environments. This study empirically analyzes this issue by designing nine training-evaluation scenarios. using three distinct datasets: academic (Set A), public internet (Set B), and directly captured unprocessed images (Set C). The results indicate that models trained on Set A experienced a 26% drop in accuracy when evaluated on Set C. In contrast, models trained on Set C maintained stable performance (F1 Score ≥ 0.83) when tested on Set A and Set B. These findings underscore the critical impact of visual discrepancies on model generalization and highlight the necessity of training with unprocessed real-world images to address domain bias for reliable AI-based gemstone identification systems.