Scholarlink

<p>Early and accurate diagnosis of leaf diseases is very important for the protection of crop yield, but most farmers still rely on manual visual inspection. This study developed a&#xa0;YOLOv12-based object detection model for the automatic multiple classification of fruit leaf diseases. A&#xa0;custom dataset of apple, tomato, cucumber, and grape leaves was built that included both healthy foliage and a&#xa0;number of common diseases with bounding box annotations. The final YOLOv12 model was trained for 100 epochs via optimized augmentation and hyperparameters. On the validation set, it reached a&#xa0;precision of 0.95, a recall of 0.93, an mAP@0.5 of 0.94, and an mAP@0.5:0.95 of 0.85, with similar performance on the held-out test set. This indicates accurate localization and classification regardless of the size of the objects and the illumination conditions. Confusion matrix analysis revealed that substances of the most frequently occurring errors are related to heavily occluded or very small leaves, which indicates directions for the expansion of a&#xa0;future dataset. Overall, the proposed system is considered a&#xa0;strong backbone for real-time and field-deployable leaf disease monitoring systems and decision support tools.</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

Real-Time Multi-Fruit Disease Detection Using YOLOv12 and a Streamlit Web Application

  • 期刊论文

  • 发表日期：

    2026年8月5日

查看全文
  • Rahul Neware,
  • Shashi Dahiya,
  • Sudeep Marwaha,
  • Shubham Singh

摘要

Early and accurate diagnosis of leaf diseases is very important for the protection of crop yield, but most farmers still rely on manual visual inspection. This study developed a YOLOv12-based object detection model for the automatic multiple classification of fruit leaf diseases. A custom dataset of apple, tomato, cucumber, and grape leaves was built that included both healthy foliage and a number of common diseases with bounding box annotations. The final YOLOv12 model was trained for 100 epochs via optimized augmentation and hyperparameters. On the validation set, it reached a precision of 0.95, a recall of 0.93, an mAP@0.5 of 0.94, and an mAP@0.5:0.95 of 0.85, with similar performance on the held-out test set. This indicates accurate localization and classification regardless of the size of the objects and the illumination conditions. Confusion matrix analysis revealed that substances of the most frequently occurring errors are related to heavily occluded or very small leaves, which indicates directions for the expansion of a future dataset. Overall, the proposed system is considered a strong backbone for real-time and field-deployable leaf disease monitoring systems and decision support tools.