Visual big data mining: toward next-generation multi-label image annotation and retrieval using Quantum Firefly optimization

Abstract This study presents a Quantum-Enhanced Firefly Algorithm (QFA)-based multi-level image annotation framework that integrates advanced Otsu thresholding, region-based feature extraction, and Bayesian multi-label classification. Images are segmented into meaningful regions using QFA to fine-tune multi-threshold Otsu segmentation, overcoming limitations of traditional Firefly Algorithm (FA) such as premature convergence and local optima. From each segmented region (blob), a 12-dimensional feature vector is extracted, capturing both color (Lab color moments) and shape (area, boundary length, convexity) properties, providing robust representations for annotation. The QFA enhances segmentation precision and efficiency through quantum-inspired probabilistic search, which allows non-local jumps and broader exploration of the threshold space, overcoming the local search limitations of traditional FA while ensuring compact and homogeneous regions and preserving edges. Unlike conventional full-image prediction methods, the proposed framework performs region-wise annotation, enabling localized labeling by associating semantic concepts with specific image regions. This region-level semantic modeling improves annotation accuracy by capturing intra-label diversity and strengthening inter-label discrimination. Evaluated on Corel A and Corel B datasets, the proposed framework achieves superior segmentation (Dice = 0.84, Jaccard = 0.70), annotation accuracy (F1-score = 0.80, mAP = 0.84), and label ranking performance (LRAP = 0.87, NDCG = 0.89) compared to traditional FA and classical Otsu methods, demonstrating its robustness for complex, multi-label image annotation tasks.