Scholarlink

The integration of Artificial Intelligence (AI) into agriculture offers substantial advantages, including increased productivity and improved product quality. Tomatoes, as a widely cultivated and consumed crop, require accurate counting and ripeness classification to streamline harvesting, estimate yields, and optimize cultivation strategies during the harvest phase. This research adopts the YOLOv11 algorithm to classify tomatoes into three ripeness categories: fully ripe, partially ripe, and green. The dataset comprises 11,267 annotated tomato instances extracted from 904 images, which were randomly divided into training (70%), validation (15%), and test (15%) sets for model training and evaluation. The YOLOv11 model achieved mAP50 scores of 0.888, 0.860, and 0.931 for fully ripe, partially ripe, and green categories, respectively. To enhance detection accuracy, especially for partially occluded or distant tomatoes, the Slicing Aided Hyper Inference (SAHI) algorithm was incorporated with YOLOv11. This hybrid approach significantly improved classification and counting performance, achieving an overall accuracy of 95% on the test set, compared to 84% when YOLOv11 was used independently.

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

An Approach for Enhancing Ripe Tomato Classification and Counting by Using Deep Learning and SAHI Image Processing

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nguyen Minh Khiem,
  • Tran Van Thanh,
  • Nguyen Thanh Hai,
  • Le Huynh Quoc Bao,
  • Nguyen Thai Nghe

摘要

The integration of Artificial Intelligence (AI) into agriculture offers substantial advantages, including increased productivity and improved product quality. Tomatoes, as a widely cultivated and consumed crop, require accurate counting and ripeness classification to streamline harvesting, estimate yields, and optimize cultivation strategies during the harvest phase. This research adopts the YOLOv11 algorithm to classify tomatoes into three ripeness categories: fully ripe, partially ripe, and green. The dataset comprises 11,267 annotated tomato instances extracted from 904 images, which were randomly divided into training (70%), validation (15%), and test (15%) sets for model training and evaluation. The YOLOv11 model achieved mAP50 scores of 0.888, 0.860, and 0.931 for fully ripe, partially ripe, and green categories, respectively. To enhance detection accuracy, especially for partially occluded or distant tomatoes, the Slicing Aided Hyper Inference (SAHI) algorithm was incorporated with YOLOv11. This hybrid approach significantly improved classification and counting performance, achieving an overall accuracy of 95% on the test set, compared to 84% when YOLOv11 was used independently.