Deep Learning-Based Classification of Maize Grain Defects Using Convolutional Neural Network

Abstract

The automated classification of post-harvest maize kernels is critical for ensuring agricultural economic value and mitigating food safety risks, particularly mycotoxin contamination from fungal infections. This study evaluates the efficacy of a deep Convolutional Neural Network, specifically the ResNet-50 architecture, for categorizing standard RGB images of maize kernels into three primary quality tiers: pure, broken, and fungal. A publicly available dataset was curated and refined to exclude minor physiological anomalies (silk cut), focusing the model entirely on the most critical structural and pathological defects. To evaluate deployability on standard, accessible hardware, the model was trained locally on a central processing unit (CPU) using a 60:20:20 data split for training, validation, and testing. The ResNet-50 model achieved a robust overall accuracy of 88% on an independent test set of 3,396 images. It demonstrated exceptional diagnostic performance for intact, healthy grain with a recall of 92.7%. Despite the well-documented computer vision challenge of morphological overlap between mechanically fractured and intact kernels, the network maintained competitive detection rates for broken (82.8% recall) and fungal-infected (81.2% recall) grain. These results indicate that standard residual architectures, operating without computationally heavy pre-segmentation algorithms or specialized hardware, offer a highly efficient and scalable baseline for real-time agricultural quality control

.