FVOR-YOLO: A Real-Time Model for Fruits and Vegetables Detection in Complex Supermarket Self-Checkout Environments

Recent advances in Computer Vision (CV) and Artificial Intelligence (AI) have rendered automated Object Detection (OD) an indispensable component of modern supermarket self-checkout (SCO) systems, particularly for barcode-less loose items such as fruits and vegetables. However, the effective deployment of these systems presents a significant technical challenge, which is to simultaneously achieve high accuracy and real-time inference speed on resource-constrained edge hardware. To address this challenge, this paper introduces FVOR-YOLO, a novel detection architecture derived from the YOLOv12 framework. The proposed model incorporates a lightweight MobileViTv2 module to replace key components of the original YOLOv12 backbone. This approach not only enhances global modelling capabilities but also preserves the local inductive biases of Convolutional Neural Networks (CNNs), resulting in a powerful yet lightweight structure. Furthermore, FVOR-YOLO integrates a parallel attention module, PCS (Parallel Channel and Spatial Group-wise Enhance), to selectively amplify salient CA and SGE features, thereby enhancing feature discriminability for OD. To address the problem of limited datasets, a bespoke fruit and vegetables image dataset is built to emulate complex supermarket SCO environments. The proposed model’s efficacy is confirmed through empirical validation. On a high-performance NVIDIA RTX 5080 GPU, the model achieves a mAP@[0.5:0.95] of 66.1%. This represents a significant 4.2% performance gain over the YOLOv12 baseline. The high accuracy is delivered alongside a rapid inference speed of 123.1 FPS. Furthermore, when deployed on the NVIDIA Jetson AGX Orin edge device, the model maintains its efficacy. The model achieves a mAP@[0.5:0.95] of 63.5%, representing a 4.3% improvement over YOLOv12, and delivers a real-time inference speed of 74.1 FPS. Collectively, these results validate FVOR-YOLO as an efficient and practical solution, particularly for edge AI applications in retail. The model is well-suited for advanced SCO systems that operate with limited computational resources.