Scholarlink

<p>Real-time monitoring of fruit growth in natural environments is an essential component of precision agriculture, encompassing disease control and identification. Existing detection algorithms face challenges such as a&#xa0;large number of parameters, thereby hindering their deployment on embedded devices. To address the internal complexity of the model, a&#xa0;lightweight detection algorithm, BD-YOLOv8, has been developed. This algorithm utilizes three collaborative modules to achieve precise classification and localization of diseases. Firstly, a&#xa0;novel image edge feature extraction module (IEFEM) was proposed. This enables the model to better extract image content by combining Sobel convolution with spatial information convolution. Moreover, an innovative detail-capture block (DC_Block) was introduced to improve the model’s perceptual ability and detection accuracy. Ultimately, a&#xa0;BN-separated parameter-sharing detection head (BSPSDH) was designed. This achieves a&#xa0;lightweight construction without loss of accuracy by sharing parameters between convolutional layers. The results indicate that BD-YOLOv8 achieves an mAP50 of 98.2% and an FPS of 149.6. Additionally, there were respective reductions of 42.4% in parameters, 27.2% in GFLOPs, and 40.3% in model size.</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

BD-YOLOv8: a Lightweight Method for Real-Time Citrus Disease Detection in Precision Agriculture

  • 期刊论文

  • 发表日期：

    2025年6月12日

查看全文
  • Xi Chen

摘要

Real-time monitoring of fruit growth in natural environments is an essential component of precision agriculture, encompassing disease control and identification. Existing detection algorithms face challenges such as a large number of parameters, thereby hindering their deployment on embedded devices. To address the internal complexity of the model, a lightweight detection algorithm, BD-YOLOv8, has been developed. This algorithm utilizes three collaborative modules to achieve precise classification and localization of diseases. Firstly, a novel image edge feature extraction module (IEFEM) was proposed. This enables the model to better extract image content by combining Sobel convolution with spatial information convolution. Moreover, an innovative detail-capture block (DC_Block) was introduced to improve the model’s perceptual ability and detection accuracy. Ultimately, a BN-separated parameter-sharing detection head (BSPSDH) was designed. This achieves a lightweight construction without loss of accuracy by sharing parameters between convolutional layers. The results indicate that BD-YOLOv8 achieves an mAP50 of 98.2% and an FPS of 149.6. Additionally, there were respective reductions of 42.4% in parameters, 27.2% in GFLOPs, and 40.3% in model size.