| 刘书婷,牟怿,陈为真.基于改进YOLOv5的碎米检测数据集[J].中国粮油学报,2024,39(4):140-148 |
| 基于改进YOLOv5的碎米检测数据集 |
| Broken Rice Detection Dataset Based on Improved YOLOv5 |
| 投稿时间：2023-04-12 修订日期：2023-07-15 |
| DOI： |
| 中文关键词: 数据集 深度学习 碎米检测 YOLOv5 |
| 英文关键词:Datasets Deep learning Broken rice detection YOLOv5 |
| 基金项目:湖北省教育厅，基于动态数据驱动的粮情智能预测模型研究（B2020061） |
|
| 摘要点击次数: 71 |
| 全文下载次数: 0 |
| 中文摘要: |
| 碎米检测是评估大米品质的重要环节，传统的碎米检测是由人工挑选完成的，这种方式耗时费力，误差率高，而且公开可用的碎米检测数据集并不多。为解决该问题，本文创建了一个大米碎米数据集，该数据集共由2435张图片和对应标签文件组成，其中包含3种类别；并提出了一个改进的YOLOv5碎米检测模型，该模型引入ShuffleNetv2轻量化结构作为特征提取结构，大大减少了模型的参数量，在此基础上，引入了BiFPN结构作为特征融合结构，使用α_IoU作为回归框损失对损失函数进行改进。实验表明，改进之后的模型精度可达98.9%，比原YOLOv5高0.3%，参数量和计算量也比原模型减少了85%以上，其中精度相比于YOLOv3、SSD、RestinaNet、FasterRCNN分别高了0.4%、33.3%、27.9%、27.2%。相关数据集将在https://github.com/THFrag/broken-rice-detection上提供。 |
| 英文摘要: |
| Broken rice detection is an important part of evaluating the quality of rice. The traditional broken rice detection is manually selected, which is time-consuming and laborious with high error rate. To solve this problem, this paper created a broken rice dataset. The dataset consists of 2435 images and corresponding label files, which contains three categories. In this model, the ShuffleNetv2 lightweight structure was introduced as the feature extraction structure, which greatly reduces the number of parameters of the model. On the basis, the BiFPN structure was introduced as the feature fusion structure, and α_IoU was used as the regression box loss to improve the loss function. Experiments shown that the accuracy of the improved model can reach 98.9%, which is 0.3% higher than that of the original YOLOv5, and the number of parameters and calculation are also reduced by more than 85% than that of the original model. Compared with YOLOv3, SSD, RestinaNet and FasterRCNN, the improved model is higher 0.4%, 33.3%, 27.9% and 27.2% in accuracy, respectively. The relevant datasets will be available at https://github.com/THFrag/broken-rice-detection. |
| 关闭 |
|
|
|
