Maize Grain Classification Using Machine Learning

Automated maize grain classification represents a critical technological challenge in precision agriculture, addressing fundamental requirements for food quality assessment and agricultural productivity. This paper proposes a hybrid framework integrating YOLO-based object detection, a two-stage Support Vector Machine (SVM) classifier, and U<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>-Net semantic segmentation to enhance classification accuracy and efficiency. The system achieves 95% accuracy while systematically categorizing maize grains into six quality classes: healthy, broken, damaged, fungus-affected, organic, and inorganic, with F1-scores between 0.93-0.98. An initial classification stage identifies primary grain groups—good grains, bad grains, and foreign materials—which are then refined through detailed sub-classification using advanced feature extraction techniques. To ensure robustness, a custom high-resolution dataset was developed using Segment Anything Model (SAM) grain isolation and Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) super-resolution. The research implements a Gradio-based interactive frontend, enabling real-time agricultural quality assessment and demonstrating the framework’s practical utility. Experimental results demonstrate 13-second processing time per image, with U<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>-Net segmentation requiring 8 seconds and classification operations 5 seconds, establishing a practical balance between accuracy and computational efficiency.