Machine vision-based detection of appearance defects in rice seeds

Screening rice seeds prior to sowing is essential for ensuring optimal crop yield. This study introduces a machine vision-based approach for detecting appearance defects, speckling, dehulling and cracking, in Wuyou Rice No. 1. Seed images were captured using a visible light camera, with a secondary segmentation technique employed to extract speckled areas. Dehulled seeds were identified through a support vector machine (SVM) utilising greyscale features. Cracks are detected using another SVM model, which processes features such as the lengths of line segments in the seed’s edge image and their Euclidean distance from the seed’s centroid. Comparative testing with alternative models reveals that, for the independent, time-separated test set, the spot detection achieved an R 2 of 0.81 and a classification accuracy of 88.7%. The RBF-kernel SVM demonstrated the best performance in dehulling and cracking detection, achieving 100% and 95.7% accuracy, respectively, surpassing CNN and BP neural networks. Receiver operating characteristic (ROC) analysis of the cracking detection model indicates strong discriminatory power (AUC > 0.9). This method provides rapid, accurate and automated seed quality assessment, supporting machine-vision-based grading. Although the current validation was based on a single variety, future research will assess the possibily to extend the method across different varieties to address variability in appearance.