BCSM-YOLO: An Improved Product Package Recognition Algorithm for Automated Retail Stores Based on YOLOv11

With the rapid growth of automated retail and smart supermarkets, commodity package recognition faces challenges like complex backgrounds, multi-scale targets, and dense occlusion. To address YOLOv11’s limitations in supermarket scenarios, such as missed small targets and low positioning accuracy, this paper proposes BCSM-YOLO, an improved algorithm based on YOLOv11. Firstly, introducing the Space-to-Depth Convolution (SPD-Conv) can maximize the preservation of detailed information such as commodity texture and shape in the downsampling stage, which provides a rich information base for the subsequent feature extraction. Then, the Convolutional Block Attention Module (CBAM) screens the processed data, adaptively focuses on the key regions, and suppresses the background interference. Subsequently, the Bidirectional Feature Pyramid Network (BiFPN) performs cross-layer bi-directional fusion of the multi-scale features processed by the CBAM to ensure that the feature information at different scales fully interacts and effectively solves the multi-scale target recognition challenges. Finally, based on the Minimum Point Distance-IoU (MPD-IoU) loss function, the alignment relationship between the bounding box output from BiFPN is precisely optimized to improve bounding box regression accuracy, and the key innovation of BCSM-YOLO is its unique combination of these advanced technologies that work together to address the specific challenges of the supermarket environment. According to the experimental results, BCSM-YOLO improved precision, recall, mAP@0.5, and mAP@0.5-0.95 by 5.5%, 6.2%, 6%, and 5.8%, respectively, under identical hyperparameter and training dataset (Ultralytics SKU-110K). Robustness tests confirm its strong anti-interference ability, demonstrating its effectiveness for detection and identification of supermarket goods.