Non-destructive identification of maize kernel varieties using a lightweight weakly supervised machine vision framework

Introduction Reliable identification of maize kernel varieties is important for seed-lot management, breeding, and quality control. However, subtle inter-varietal phenotypic differences make rapid, non-destructive identification challenging. This study aimed to develop an accurate and computationally efficient identification method using only image-level labels. Methods We developed Decoupled Refinement with Adaptive Hyperparameter Optimization (DRAHO), which integrates a MobileNetV4-Conv-Medium backbone, a decoupled Global -Locator -Reviewer (GLR) network, and Elite Adaptive Rime Optimization (EARIME). The GLR network performs global recognition, annotation-free localization, and attention-guided feature refinement without explicit image cropping. EARIME optimizes the learning rate and loss weights using a fixed proxy task based on validation performance. DRAHO was trained on 10,000 individual kernel images from ten maize varieties and evaluated on a held-out test set. Results DRAHO achieved 98.30% accuracy, exceeding Progressive Multi-Granularity (PMG) by 2.40 percentage points, with 7.16 million parameters and 1.65 GFLOPs. The proxy configuration reduced the nominal sample -epoch workload per candidate by 96.25% compared with full-data training for 100 epochs. After independent dataset-specific retraining, DRAHO achieved 97.50% accuracy on a public maize dataset and 98.26% accuracy on a public wolfberry dataset. Discussion These results demonstrate that DRAHO combines weakly supervised feature refinement with efficient hyperparameter selection. Under controlled imaging conditions, it provides a practical approach to non-destructive maize kernel variety identification with modest computational requirements.