Research on Crack Detection for Tobacco Stems with Improved YOLOv5

To achieve accurate detection of cracks in blasting tobacco stems and facilitate the subsequent implementation of intelligent grading, this study proposes an improved model for crack detection based on YOLO, a target detection algorithm. The proposed model integrates the Swin Transformer module into the YOLOv5 detection model, which decreases the number of parameters while enhancing crack features extraction. Additionally, the CBAM attention mechanism is introduced to address the model's limitations in capturing target feature information during tobacco stem crack detection. The NWD loss function is employed instead of CIoU Loss as the bounding box loss function, accelerating network convergence and improving detection accuracy. Furthermore, the Decoupled Head replaces the detection head in YOLOv5s, separating the classification task from the regression task. Compared with the original YOLOv5s, the optimized model improves the Average Precision (mAP) by 4.5% and increases computational speed by 5.3 frames per second (fps). On a locally sourced tobacco stem dataset, the final model achieves a detection mAP of 85.2% and a recall (R) of 78.3%, with a detection speed of 38.9 fps on a local computer. These results offer valuable insights for advancing crack identification as well as real-time online sorting and detection technologies for blasting tobacco stems.