Shelf Image Analysis Using YOLO11 and YOLOv12 for Product Recognition

Effective product identification and inventory tracking by analyzing the shelves are vital for streamlining operations, better profitability, and improved customer experience in the retail supermarkets. The aim of this study is to identify and count the number of products on the shelves using annotated images of 27 product class labels in Grozi-3.2k dataset, which has a diverse set of images in varied angles and layouts of shelves. The latest models of the You Only Look Once (YOLO) algorithm, such as YOLO11 and YOLOv12 in Small (S) and Medium (M) variants were used for training the dataset and Nvidia A100 Graphical Processing Units (GPUs) in Google Colab was used as the hardware configuration for training this model. It was observed that the YOLO11-S model achieved better results when compared to other models with 0.7962 for mAP@0.5, 0.79 for precision, and a test accuracy of$\mathbf{9 4. 8 9 \%}$for IoU@0.5, which is closely followed by the YOLOv12-S model, whereas the$M$variants of both models lagged compared to the S variants in performance. This work highlights the YOLO model's potential to be successful in real-time retail inventory management at the supermarkets when compared to manual identification and counting of the products. This study lays the foundation for further advancements in the field of AI-driven systems for retail inventory management.