Introduction: Corn (Zea mays L.) is a strategic crop for global food security, feed, and biofuel. Genetic purity of hybrid seeds is vital for high yield and stress resistance. However, different corn varieties often look very similar. Subtle differences in color, texture, shape, and size make traditional identification difficult. Manual inspection is time-consuming, subjective, error-prone, and not scalable. Thus, an automated, non-destructive, and accurate system is needed. Deep learning, especially CNNs, is a powerful tool for agricultural image recognition. Mobile-optimized networks like MobileNet and NASNetMobile offer a good balance between accuracy and efficiency. They are suitable for smartphones and embedded systems. This study compares six deep learning architectures for non-destructive identification of five Iranian corn seed varieties.

Materials and Methods: Six deep learning architectures were evaluated including base MobileNet, DenseNet121, MobileNetV2, MobileNetV3Small, MobileNetV3Large, and NASNetMobile. All models were pre-trained on ImageNet and fine-tuned for corn seed classification. A total of 9,374 high-resolution images of five corn seed classes (KSC201, KSC260, KSC400, KSC703, KSC705) were acquired under controlled conditions. The dataset was split into training (6114 images, 65.2%), validation (1875 images, 20%), and test (1385 images, 14.8%) sets. Models were trained using the Adam optimizer (learning rate 0.0001) for 50 epochs with categorical cross-entropy loss. Data augmentation (rotation, shift, zoom, flip) was applied to enhance generalization. Performance was evaluated using accuracy, precision, recall, specificity, F1-score, ROC curves, AUC, confusion matrices, and training time.

Results and Discussion: MobileNet achieved the best overall performance with 97.04% accuracy, 96.66% precision, 96.70% recall, 99.27% specificity, and 96.67% F1-score. DenseNet121 ranked second with 95.96% accuracy and 95.41% F1-score. MobileNetV3Small (95.02% accuracy, 94.20% F1-score) outperformed MobileNetV3Large (93.43% accuracy, 92.82% F1-score). MobileNetV2 achieved 93.14% accuracy and 92.71% F1-score, while NASNetMobile recorded the weakest performance with 87.08% accuracy and 86.22% F1-score. All models showed high specificity ranging from 96.77% to 99.27%. Class-wise analysis revealed that KSC260 was the most stable class with F1-scores of 98.01% for MobileNet, 98.15% for MobileNetV3Small, 98.13% for DenseNet121, and 96.9% for MobileNetV2. KSC703 also performed well with F1-scores of 98.42% for MobileNet and 98.27% for DenseNet121. In contrast, KSC705 was the most challenging class; MobileNetV2 showed extreme behavior on KSC705 with 99.24% recall but only 77.74% precision, resulting in a low 87.19% F1-score. Persistent misclassifications occurred between class pairs KSC201/KSC705 and KSC400/KSC705 across all models, indicating significant feature overlap. ROC curve analysis showed that MobileNet, DenseNet121, MobileNetV3Small, and MobileNetV3Large achieved an ideal AUC of 1.00, while MobileNetV2 and NASNetMobile recorded excellent AUC values of 0.99 and 0.98, respectively. Regarding computational efficiency, training time varied significantly: MobileNet required 108.32 minutes (highest), DenseNet121 required 104.66 minutes, while MobileNetV3Small required only 62.28 minutes, offering a 46-minute reduction compared to MobileNet while maintaining 95.02% accuracy. MobileNetV2 (60.77 min), MobileNetV3Large (63.96 min), and NASNetMobile (64.04 min) showed moderate efficiency. There is a clear trade-off between accuracy and computational efficiency: MobileNet provides maximum accuracy (97.04%) but requires the most training time (108.32 min), whereas MobileNetV3Small offers a balanced solution with 95.02% accuracy and 62.28 min training time. Therefore, for maximum absolute accuracy without computational constraints, MobileNet is the top choice; for a balanced trade-off between accuracy and efficiency, DenseNet121 is suitable; and for resource-constrained environments such as mobile devices, MobileNetV3Small is the best option.

Conclusion: This study demonstrates that mobile-optimized deep learning architectures, particularly base MobileNet and MobileNetV3Small, possess high potential for accurate, non-destructive, and efficient classification of corn seed varieties. These findings provide a practical reference for researchers and engineers in automated agricultural product classification, offering clear architecture selection guidelines based on specific application constraints.