Deep Learning and Computer Vision for Maize Seed Classification: A Review of Current Methods and Future Directions

Abstract

Seed purity is a key indicator of maize seed qual-ity and vital for food security. This study presents a non-destructive, high-throughput system combining RGB imaging, hyperspectral data, near-infrared sensing, and biopecle analysis to separate pure and mixed seeds. Deep-learning models such as Swin Transformer, EfficientNetV2, and self-supervised DINOv2 features are employed, with seed segmentation performed using the Segment Anything Model. The pipeline includes dataset collection, modality-specific preprocessing, multimodal fusion, and comprehensive evaluation with ablation studies. The goal is to achieve over 95% purity-classification accuracy. Key contri-butions include a practical multimodal framework and a hybrid supervised, self-supervised training strategy.