您好，欢迎访问湖南省农业科学院 机构知识库！

Improved CSW-YOLO Model for Bitter Melon Phenotype Detection

文献类型： 外文期刊

作者： Xu, Haobin 1 ; Zhang, Xianhua 1 ; Shen, Weilin 1 ; Lin, Zhiqiang 4 ; Liu, Shuang 1 ; Jia, Qi 3 ; Li, Honglong 5 ; Zheng, Jingyuan 2 ; Zhong, Fenglin 1 ;

作者机构： 1.Fujian Agr & Forestry Univ, Coll Hort, Fuzhou 350002, Peoples R China

2.Hunan Acad Agr Sci, Inst Vegetables, Changsha 410125, Peoples R China

3.Jiuquan Acad Agr Sci, Jiuquan 735099, Peoples R China

4.Fujian Agr Machinery Extens Stn, Fuzhou 350002, Peoples R China

5.Fujian Tianmei Seed Ind Technol Co, Fuzhou 350109, Peoples R China

关键词： bitter melon; phenotypic detection; deep learning; CSW-YOLO

期刊名称：PLANTS-BASEL （ 影响因子：4.1； 五年影响因子：4.5 ）

ISSN： 2223-7747

年卷期： 2024 年 13 卷 23 期

页码：

收录情况： SCI

摘要： As a crop with significant medicinal value and nutritional components, the market demand for bitter melon continues to grow. The diversity of bitter melon shapes has a direct impact on its market acceptance and consumer preferences, making precise identification of bitter melon germplasm resources crucial for breeding work. To address the limitations of time-consuming and less accurate traditional manual identification methods, there is a need to enhance the automation and intelligence of bitter melon phenotype detection. This study developed a bitter melon phenotype detection model named CSW-YOLO. By incorporating the ConvNeXt V2 module to replace the backbone network of YOLOv8, the model's focus on critical target features is enhanced. Additionally, the SimAM attention mechanism was introduced to compute attention weights for neurons without increasing the parameter count, further enhancing the model's recognition accuracy. Finally, WIoUv3 was introduced as the bounding box loss function to improve the model's convergence speed and positioning capabilities. The model was trained and tested on a bitter melon image dataset, achieving a precision of 94.6%, a recall of 80.6%, a mAP50 of 96.7%, and an F1 score of 87.04%. These results represent improvements of 8.5%, 0.4%, 11.1%, and 4% in precision, recall, mAP50, and F1 score, respectively, over the original YOLOv8 model. Furthermore, the effectiveness of the improvements was validated through heatmap analysis and ablation experiments, demonstrating that the CSW-YOLO model can more accurately focus on target features, reduce false detection rates, and enhance generalization capabilities. Comparative tests with various mainstream deep learning models also proved the superior performance of CSW-YOLO in bitter melon phenotype detection tasks. This research provides an accurate and reliable method for bitter melon phenotype identification and also offers technical support for the visual detection technologies of other agricultural products.

  • 相关文献
作者其他论文 更多>>
置顶
购物车
反馈

© 京ICP备09089781号-29 主办单位：湖南省农业科学院

学术资源： 中国农业科学院 农业专业知识服务系统 农科联盟资源共建共享平台 中国农业科技文献与信息服务平台 中国农业期刊集成服务平台