Scholarlink

<p>Mangosteen grading is essential for maintaining quality standards in both local and export markets. Traditional manual grading, based on visual inspection, is time-consuming and inconsistent. This paper proposes a multi-view regression-based model using convolutional neural networks (CNN) to automate the grading process. Methodologically, the proposed architecture employs two shared CNN-backbones to extract spatial features from six views, where one backbone processes the top and bottom views, while another processes the four side views. The extracted features are aggregated into a regressor to predict a continuous quality score (0–1). This score is then mathematically mapped to a discrete grade class via a proximity function, flexibly accommodating different market standards without structural changes. Trained on datasets from three trading markets, the model achieves grading accuracies of 100%,&#xa0; 95%,&#xa0; and 99% for three, seven, and eight class datasets, respectively.</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

Mangosteen grading using image regression under multiple views

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月24日

查看全文
  • Worapan Kusakunniran,
  • Kittinun Aukkapinyo,
  • Kittikhun Thongkanchorn,
  • Pimpinan Somsong,
  • Pimsiri Tiyayon,
  • Sai Thu Ya Aung,
  • Yu Nandar Aung,
  • Thirada Suesat

摘要

Mangosteen grading is essential for maintaining quality standards in both local and export markets. Traditional manual grading, based on visual inspection, is time-consuming and inconsistent. This paper proposes a multi-view regression-based model using convolutional neural networks (CNN) to automate the grading process. Methodologically, the proposed architecture employs two shared CNN-backbones to extract spatial features from six views, where one backbone processes the top and bottom views, while another processes the four side views. The extracted features are aggregated into a regressor to predict a continuous quality score (0–1). This score is then mathematically mapped to a discrete grade class via a proximity function, flexibly accommodating different market standards without structural changes. Trained on datasets from three trading markets, the model achieves grading accuracies of 100%,  95%,  and 99% for three, seven, and eight class datasets, respectively.