Mangosteen grading using image regression under multiple views

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

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.