Research on Cigarette Recognition based on Improved YOLOv5

Abstract

To address the critical challenges of misclassification in automated strip tobacco sorting and inefficiency in manual verification within tobacco logistics centers, this study presents an enhanced YOLOv5s-based framework that optimizes real-time performance and recognition accuracy through systematic architectural improvements. First, a Coordinate Attention (CA) mechanism is strategically integrated into the backbone network to enhance discriminative feature extraction, enabling precise localization of tobacco strip targets. Subsequently, the conventional nearest-neighbor upsampling operator is replaced with the Content-Aware ReAssembly of FEatures (CARAFE) module, effectively expanding the receptive field while maintaining computational efficiency. Furthermore, a Ghost convolution-based lightweight redesign is implemented in the feature extraction layers, achieving a 45.8% reduction in model parameters without compromising detection capability. Complementing these algorithmic advancements, we construct a dedicated tobacco image acquisition system under industrial operating conditions, compiling a domain-specific dataset containing 12,800 annotated instances of strip tobacco with morphological variations. Experimental validation demonstrates that the optimized model achieves a mean average precision (mAP@0.5) of 99.3% and maintains 99.9% classification accuracy on the error correction testbed, while operating at 48 FPS on an NVIDIA Jetson Xavier NX edge device. These quantitative results confirm the framework's capability to fulfill the stringent requirements of high-speed industrial sorting systems, achieving a 21.6% improvement in processing throughput compared to baseline YOLOv5s while sustaining sub-millisecond latency thresholds.