The quantification of peck damage in rice kernels using visual features of the high magnification images

Pecky or pin damage in rice caused by stink bugs is a key quality factor in determining storage stability, market approval, and end-use. Peck-infected kernels exhibit chalkiness, yellowing, and discolored spots in black, brown, red, or other colors. The varied damage patterns suggest that stink bugs may not be the sole cause, and a comprehensive evaluation could clarify these variations. However, detailed quality assessment for classifying peck damage in milled rice remains unexplored due to limited datasets and deep learning studies. In this study, a machine vision system is developed to construct a dataset of 2381 high-resolution ( 24 Mega − pixel ) and high-magnification ( 3.9 µm / pixel ) images of peck-damaged rice kernels. An image segmentation model is trained for pixel-wise damage classification. An unsupervised segmentation approach based on the high feature descriptiveness of Convolutional Neural Networks is investigated to address the labor-intensive process of creating ground truth masks for images. Unsupervised segmentation predictions are used to calculate parameters such as peck damage area, centroid location, and scatter along the rice grain’s length, ultimately classifying the damages into four variations. The final classification utilizes geometric data from the unsupervised segmentation model through the K-means approach. The effectiveness of the unsupervised segmentation model is validated with a dice coefficient of 0.9254 on a reference dataset, and visual inspection of the segmented images confirms successful segmentation of the damaged areas. Overall, this research enhances the understanding of variations in peck damaged milled rice grains and helps to identify the underlying causes of these variations. The framework has direct implications for rice quality assessment, automated grading, and decision-support tools for post-harvest processing. • Developed a dataset of 2381 high-resolution rice kernel images • Segmentation algorithm classifies pecky damage based on geometric features • Unsupervised learning approach eliminates the need for extensive manual labelling • Peck damage is classified into four variations using K-means clustering • Achieved 0.9254 Dice coefficient for unsupervised segmentation validation