Unified parcel attribute recognition: an efficient edge-based system using YOLOv12-BiFPN-ResCBAM for barcode, dimensioning, and damage detection

Automated parcel processing is pivotal for enhancing efficiency within the logistics industry; however, its widespread adoption is impeded by the prohibitive costs of specialized hardware and the computational constraints associated with edge deployment. To address these limitations, this study presents a unified, cost-effective framework capable of consolidating three critical functions—barcode recognition, volumetric dimensioning, and damage detection—into a single device. We propose an enhanced architecture, designated as YOLOv12-BiFPN-ResCBAM, which incorporates a Bi-directional Feature Pyramid Network (BiFPN) and a Residual Convolutional Block Attention Module (ResCBAM) to optimize the trade-off between detection precision and inference latency. Experimental validation demonstrates that the system achieves a processing throughput of 44 FPS and a dimensional error rate of approximately 2.5%, performance metrics that align with established industrial standards. Notably, the optimized small variant attained an mAP50:95 of 95.68% (+1.5% over baseline) while concurrently reducing the parameter count by 11% and model size by 10.7%. These findings substantiate the proposed system as a scalable, high-precision, and computationally efficient alternative to conventional logistics infrastructure.