부산대학교 도서관

This paper presents a novel deep learning-based method for multi-class price tag detection in shelf images that capture the full shelf height. In the proposed method, the YOLOv8 price tag detection model is used to localize and classify the price tags in an image. We develop a new augmentation method where some of the existing price tags in a shelf image are duplicated and shifted horizontally. In addition, price tag crops are used as templates to generate synthetic price tag images. In order to solve the price tag class imbalance problem, we also develop a method for balancing the ratio of different price tag classes in shelf images using the aforementioned synthetic price tags. Finally, we train the price tag detection model using balanced data. Experimental results show that the five-class price tag detection model trained on unbalanced data achieves mean average precision ($m$AP) of 73.1%, while training on balanced data helps to boost the result of price tag detection up to 93.9% $m$AP.