Tlb‑yolo: a rapid and efcient real‑time algorithm for box-type classification and barcode recognition on the moving conveying and sorting systems

Abstract In recent years, camera-based vision sensors utilizing YOLO neural network algorithms have increasingly replaced traditional sensors such as photoelectric sensors. Due to their advantages in cost and efficiency, applications in object classification and barcode detection have become more widespread in industrial and logistics sectors. This study addresses the real-time items classification and barcode detection by proposing the Transmission Line Barcode YOLO (TLB-YOLO) model. We made improvements to the YOLOv8 model by introducing several new components. We integrated the Coordinate Attention (CA) mechanism into the backbone network to improve the model's sensitivity to object locations. The Wise-IoU loss function was employed to enhance localization accuracy, while the GSConv (Grouped Shuffle Convolution) and Slim Neck architecture were incorporated to boost detection accuracy and speed. The proposed model, with 3.8 million parameters and 8.5 GFLOPs, was trained on the COCO dataset, achieving an mAP0.5 of 68.1% and an mAP0.95 of 44.2%, which represent improvements of 7.9% and 5.2% over YOLOv8n, respectively. It attained a frame rate of 153.8 FPS. When retrained on a custom dataset incorporating synthetic data from Omniverse, the model demonstrated over 90% accuracy in detecting cardboard boxes, plastic containers, and barcodes. To resolve model export issues caused by the dynamic pooling layer, an equivalent code substitution was applied, enhancing inference speed. Testing on a Jetson Nano development board, with each experiment phase repeated 50 times, showed that the TLB-YOLO model achieved 100% detection accuracy for plastic containers side scanning, cardboard box top scanning, and barcode verification. The model’s detection accuracy and inference speed fully satisfy real-time detection requirements in practical scenarios.