StockSense-R2: Robust and Cost-Aware Deep Learning for Retail Shelf-Void Detection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Efficient detection of empty retail-shelf regions can support timely replenishment and reduce stockout risk. This study presents StockSense-R2, a robust and cost-aware shelf-void detection framework based on YOLOv8s. The framework combines transfer learning, controlled augmentation, confidence-threshold optimization, test-time augmentation, visual-perturbation testing, and an asymmetric operational-cost function that penalizes missed shelf voids more heavily than false alerts. The proposed model achieved a matched-box F1-score of 0.929 on the validation set and 0.913 on the independent test set, increasing to 0.915 with test-time augmentation. The test evaluation recorded an mAP@0.50 of 0.875 and an mAP@0.50:0.95 of 0.609. The model maintained near-real-time performance at 24.37 images per second with a mean inference latency of 41.03 ms per image. Robustness testing showed stable performance under darkness, low contrast, and glare, while noise, blur, and occlusion remained more challenging. These findings indicate that StockSense-R2 provides a practical approach for robust and cost-sensitive retail shelf monitoring.