A Machine Vision-Based Dual-View Online Quality Grading Method for Walnut Kernels

Abstract

To improve the efficiency and completeness of walnut kernel grading, a dual-view machine vision-based online grading method and corresponding automated sorting system were developed. Engineering-oriented criteria classified kernels into four grades based on colour and defects, while rejecting dark, damaged, mouldy, and impurity-containing samples. Dual cameras captured top and bottom surfaces, and a spatial matching strategy enabled reliable dual-view association. A fusion-based decision method determined overall quality. A lightweight object detection model, YOLO-SR, was proposed by integrating ShuffleNetV2 and C3k2_RFAConv, reducing computational load and model size by 34.9% and 30.9%, respectively, while maintaining accuracy. Deployment results showed an accuracy of 98.83% and throughput of 0.1667 kernels·s-1, with stable performance under varying loads and 120 min continuous operation. The method achieved comparable accuracy but significantly higher efficiency than manual grading, demonstrating strong potential for industrial application.