基于YOLOv5的烤烟烟叶散把程度检测算法研究

为解决烤烟烟叶散把过程中因散把不均匀导致烟叶重叠等问题，提出了一种基于YOLOv5目标检测算法的烟叶散把程度检测方法。通过对原始图像进行预处理构建烟叶散把图像数据集，在原始YOLOv5模型主干网络加入Ghost模块生成冗余特征图，在瓶颈层加入ACIN模块加强网络特征融合，同时利用烟叶松散度来评价散把程度。分别利用改进前后YOLOv5模型进行测试，结果表明:与原始模型相比，改进后YOLOv5模型在未明显增加计算量的前提下，网络参数量减少12.8%，模型大小减小12.4%，平均精确率提升0.2百分点；改进后模型与YOLOv4、Efficientdet-d0、Faster R-CNN等目标检测模型相比，平均精确率、检测速度均为最优且参数量较少。该技术可为提高烤烟烟叶分选速度和精度提供支持。 

YOLOv5-based detection algorithm for evaluating degree of bundle-loosening of flue-cured tobacco

To prevent tobacco leaves from overlapping caused by imperfect bundle-loosening, a method for detecting the loosening degree of tobacco bundles was proposed based on YOLOv5 object detection algorithm. The original images were preprocessed to build a dataset of loosened tobacco bundle images. A Ghost module was added to the backbone network of the original YOLOv5 model to generate redundant feature maps. An ACIN module was added to the bottleneck layer to enhance network feature fusion. Meantime, the loosening degree of tobacco leaves was used to evaluate the loosening degree of tobacco bundle. The YOLOv5 models before and after modification were comparatively tested, and the results showed that on the premise of no significant calculation amount addition, the modified YOLOv5 model reduced the number of network parameter and the size of model by 12.8% and 12.4% respectively, and increased average accuracy by 0.2 percentage points. Compared with object detection models, YOLOv4, Efficientdet-d0 and Faster R-CNN, the modified YOLOv5 model features the highest average accuracy and detection speed and fewer Parameters. This technology could provide a support for promoting the speed and precision of flue-cured tobacco sorting.