Abstract
The accurate and rapid evaluation of cereal grain quality is fundamental to global food security, efficient processing, and fair trade. Traditional manual inspection is constrained by subjectivity and labor intensity, prompting the development of automated, non-destructive methods. This review critically examines the technological evolution from classical machine vision systems based on handcrafted features to deep learning architectures, and finally to vision transformers. It analyzes how early approaches leveraged morphological, color, and texture descriptors under controlled conditions to achieve high classification accuracy, yet often exhibited limited robustness under domain shift. The emergence of convolutional neural networks enabled end-to-end feature learning, improving robustness for tasks such as varietal classification, defect detection, and sorting. More recently, vision transformers introduced a paradigm shift by modeling global dependencies and capturing subtle, spatially distributed patterns, offering advantages for complex diagnostic tasks, while their data hunger and computational demands remain significant bottlenecks. Beyond summarizing architectures, this review synthesizes reported evidence into an analytical perspective on accuracy–latency–memory trade-offs, the dominant failure modes that drive lab-to-line performance gaps, and the evaluation practices required to claim deployable reliability. We highlight enabling factors such as large-scale open datasets, model lightweighting for edge deployment, and explainable AI to enhance transparency and trust, and we outline future trends toward multi-modal sensing, hybrid CNN–Transformer architectures, zero-shot learning, and fully autonomous inspection-sorting systems. By framing these developments as responses to technical and practical constraints, this review provides a structured and actionable perspective on the past, present, and future of intelligent grain quality assessment.
