A Quantitative Descriptor for Chalky and Discoloration Damage in Rice Grains

Abstract

The quality assessment of chalky and discolored rice grains is often limited to either subjective methods or single-value metrics that fail to capture the finer details of these damages. Despite being detectable and rich in visual features, there is a lack of a quantitative framework for characterizing these damages within a single rice grain in terms of their severity, their precise location, and their dispersion patterns across the grain surface. To address this gap, a multifeature-based computer vision framework is proposed on a dataset of 5598 high-resolution (24 megapixels) and high-magnification (3.9 μm/pixel) images. The framework involves a two-stage process: first, the unsupervised segmentation of the images via K-means clustering; and second, the spatial, geometric, and intensity-based features are extracted from the segmentation masks to create a quantitative descriptor for chalkiness and discoloration. Distinct patterns appear for each damage type. For discolored grains, a strong negative correlation (−0.75) is observed between the damaged area and its distance from the grain’s center, which indicates that the smaller defects are often localized near the grain head. In contrast, the chalky grains exhibit a strong positive correlation (0.80) between the circularity of the damage area and its relative intensity, which indicates that a damage that is more compact tends to be visually more prominent. These results put together indicate that a multimetric approach can offer a more detailed characterization of the grain quality than a single measurement based simply on the percentage of damaged area in the grain. Quantifying the degree of discoloration and chalkiness in terms of both severity and spatial dispersion provides a more objective and physically meaningful basis for grading. This multimetric representation reduces subjectivity, minimizes ambiguity caused by coarse classifications or limited-resolution imaging, and enables finer discrimination among damage patterns. Overall, such an approach would support the development of more robust, reproducible, and transparent grading systems and improve the accuracy of rice quality assessment while bridging the gap between simple manual inspection methods and computationally expensive high-resolution techniques.