Investigating performance and key factors for real-world deployment of grain image classification using convolutional neural networks

Accurate and efficient grain quality assessment is critical for making informed decisions throughout the grain value chain. Early detection of disease enables actions to mitigate spread and further damage, and optimal batch mixing to fulfill specified quality requirements allows for maximizing value and minimizing scrapping. Vision based machine learning and deep learning approaches are gaining attention in the agricultural sector and are useful for the development of automated grain quality assessment. These techniques can reduce the current manual inspection load and are key for objective and precise analysis. Yet, the majority of prior studies are constrained to small or controlled and curated datasets. Practical challenges associated with real-world deployment and reliability are rarely addressed. That is the focus of this work. We present and demonstrate a structured approach for investigating convolutional neural networks (CNNs) and key factors influencing performance for wheat kernel classification. The objective is to determine a CNN model that ensures high and robust classification accuracy, while elucidating and explaining how different image dataset characteristics and training parameters affect performance and reliability. We use a commercial mirror-based imaging system that captures over 90% of each kernel's surface and contrast and compare model architectures, robustness, the effect on pre-processing and image resolution. Our results show similar and high overall performance for ResNet50V2 and EfficientNetV2B0 ([Formula: see text]% accuracy), but per-class analysis indicate that the smaller classes suffer from lack of representative examples, and that most classes benefit from pre-processing including downsampling whereas others benefit from higher resolution. Interactive visualizations reveal that another contributing factor is dubious annotation and multi-class belongingness. Thus, our step-by-step analysis of CNN performance underscores the need for representative data, proper pre-processing, and class-aware evaluation to ensure trustworthy deployment in wheat grain quality assessment.