Scholarlink

<p>In order to solve the problem of intelligent sorting of walnut shell kernel materials facing a low level, this study proposes a lightweight YOLO9t-DM multi-target detection model based on YOLOv9 improvement. Firstly, by replacing average convolution and partial standard convolution with Dual Convolutional Kernels (DualConv), the network complexity can be reduced and the original information of input feature graphs can be better retained. Then, we use the Intersection over Union with Minimum Points Distance (MPDIoU) as the boundary box loss function to solve the existing loss function problem and more accurate regression results. The improved model was trained and tested using the self-built Yangbi walnut material data set after breaking the shell. The results showed that the overall recognition accuracy, recall rate, Mean Average Precision (mAP) of mAP0.5 and mAP0.5<sub>:0.95</sub> reached 97.34%, 96.44%, 98.37% and 82.88% respectively. The weight Size of the model is 4.48&#xa0;MB, and the frame rate FPS is 61f/s. Compared with the improvement before, mAP0.5 and mAP0.5<sub>:0.95</sub> increased by 0.3 and 0.57 percentage points respectively, the Size decreased by 0.34&#xa0;MB, and the FPS increased by 9 units. The model outperforms Faster R-CNN, EfficientDet, and other YOLO lightweight models of the same size, with better performance for integrated applications. The model proposed in this study can improve the material detection accuracy under the premise of ensuring lightweight, which provides theoretical research and technical support for intelligent sorting of walnut shell kernel materials.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

YOLOv9t-DM: a lightweight multi-target detection method for walnut shell kernel materials

  • 期刊论文

  • 发表日期：

    2025年5月19日

查看全文
  • Guanming Yang,
  • Yingbiao Wang,
  • Xinying Li,
  • Quanxin Cui,
  • Taifeng Luo,
  • Kejing Wang

摘要

In order to solve the problem of intelligent sorting of walnut shell kernel materials facing a low level, this study proposes a lightweight YOLO9t-DM multi-target detection model based on YOLOv9 improvement. Firstly, by replacing average convolution and partial standard convolution with Dual Convolutional Kernels (DualConv), the network complexity can be reduced and the original information of input feature graphs can be better retained. Then, we use the Intersection over Union with Minimum Points Distance (MPDIoU) as the boundary box loss function to solve the existing loss function problem and more accurate regression results. The improved model was trained and tested using the self-built Yangbi walnut material data set after breaking the shell. The results showed that the overall recognition accuracy, recall rate, Mean Average Precision (mAP) of mAP0.5 and mAP0.5:0.95 reached 97.34%, 96.44%, 98.37% and 82.88% respectively. The weight Size of the model is 4.48 MB, and the frame rate FPS is 61f/s. Compared with the improvement before, mAP0.5 and mAP0.5:0.95 increased by 0.3 and 0.57 percentage points respectively, the Size decreased by 0.34 MB, and the FPS increased by 9 units. The model outperforms Faster R-CNN, EfficientDet, and other YOLO lightweight models of the same size, with better performance for integrated applications. The model proposed in this study can improve the material detection accuracy under the premise of ensuring lightweight, which provides theoretical research and technical support for intelligent sorting of walnut shell kernel materials.