Grainalyze: A Hybrid Approach for Consumer Level Assessment of Rice Quality Based on Grain Morphology

Abstract

Rice is a staple food in the Philippines, and its quality significantly affects consumer preference, market value, and grain grading standards across different regions and local agricultural markets today. However, conventional rice quality assessment commonly relies on manual inspection by trained personnel, making the process time-consuming, subjective, and often inaccessible to ordinary consumers.
This study proposes Grainalyze, a mobile-based artificial intelligence-assisted system for rice grain quality assessment using instance segmentation and image-based analysis. Image-based rice quality assessment using computer vision and machine learning techniques has been widely explored in previous studies [1], [7]. A total of 3,885 source images of rice grains were prepared and expanded to 10,135 images through preprocessing and augmentation, then annotated into four grain quality classes: whole grain, broken grain, chalky grain, and discolored grain.
The study evaluated Mask R-CNN alongside YOLOv8n-seg, YOLOv8s-seg, and U-Net using standard performance metrics, including Precision, Recall, F1 Score, mean Average Precision, and inference speed. Deep learning-based segmentation approaches have demonstrated strong performance in rice grain classification tasks [11], [19]. Results show that Mask R-CNN achieved the best overall segmentation performance, obtaining 89.89% mean Average Precision, 96.88% precision, 89.89% recall, and a 93.26% F1 Score, demonstrating the most reliable balance between segmentation accuracy and detection performance among the evaluated models.
The findings demonstrate that integrating instance segmentation with coin-based measurement calibration, rule-based broken grain detection, Logistic regression for chalky grain identification, and LAB/HSV-based discoloration analysis can provide an effective and accessible approach for automated rice grain quality assessment using smartphone-captured images.