A machine vision system for real-time grain quality classification using machine learning

Abstract

Grain quality control directly affects the efficiency of grain processing, the stability of product parameters, and the economic outcomes of storage and cleaning operations. In industrial practice, visual inspection and manual sampling remain widespread, yet these approaches are time-consuming and sensitive to human subjectivity, especially when large grain volumes must be assessed continuously. This study addresses the need for an automated, objective, and scalable solution for grain quality classification that can operate in real time under conveyor-based conditions. The purpose of the research is to develop and experimentally validate a machine vision system that classifies grain by quality and supports operational decisions for fractional separation and cleaning. An experimental test stand was built to simulate conveyor transportation of grain, enabling controlled variation of belt speed and illumination conditions. A dataset of wheat, corn, and barley was formed using laboratory image capture and manual labeling into three quality classes based on visible defects and damage severity. Image preprocessing and augmentation were applied to increase variability and improve robustness. Two supervised approaches were implemented for comparative evaluation: a deep learning image classifier and a kernel-based classifier using handcrafted visual descriptors. The experimental results demonstrate that the deep learning approach achieves higher classification accuracy, while the kernel-based classifier provides faster inference with a moderate reduction in accuracy. The most frequent misclassifications occur between adjacent quality categories, indicating the importance of borderline-class coverage and labeling consistency. Processing time measurements confirm the feasibility of real-time operation for moderate grain flow rates, with performance degradation at higher conveyor speeds due to motion-related image distortions. The scientific novelty lies in the integrated experimental assessment of classification accuracy and throughput under controlled conveyor conditions. The practical significance is the feasibility of deploying the system as a component of automated grain cleaning and separation lines.