Assessing deep learning model robustness for banana ripeness classification under varying illumination conditions

The classification of banana ripeness is critical for the external assessment of fruit quality in automated inspection systems. This study investigates the robustness of deep learning models under varying illumination conditions, utilizing a dataset of real banana images captured in multiple lighting scenarios. Models were trained and evaluated with gamma-based illumination augmentation generated through the Low-light Image Enhancement model to simulate realistic variability. Although training on pristine images yields high baseline accuracy, the models exhibit pronounced performance drops under altered illumination, highlighting significant overfitting issues. Augmentation substantially improves robustness, but the degree of improvement is architecture-dependent and does not fully mitigate vulnerabilities under extreme lighting conditions. Furthermore, the trade-offs between model complexity, inference efficiency, and generalization capacity are critically analyzed, revealing constraints for real-time and resource-limited applications. In particular, baseline accuracies reached an average of 92.5 % under ideal lighting (I0), with ViT scoring 93.13 % and InceptionV3 trailing at 89.84 %. Under moderate gamma augmentation ( Γ = 0.4 ), performance dropped by up to 11.8 percentage points (InceptionV3: from 89.84 % to 77.63 %). Training with augmented data restored accuracy by 8.3 pp on average, with InceptionV3 recovering to 91.88 % and ViT maintaining 92.90 %. Inference times ranged from 2.08 ms (ViT) to 3.89 ms (InceptionV3), demonstrating feasibility for real-time deployment. These findings emphasize the importance of incorporating environmental variability into evaluation protocols to ensure a reliable deployment of automated grading systems in practical scenarios. • Synthetic augmentation of lighting variations using the LIME model was introduced to the dataset for banana ripeness classification, aiming to evaluate and improve the model's robustness and stability under different lighting conditions. • Deep learning models show significant performance drops when lighting variations are introduced into the dataset. • The dataset, including gamma correction augmentations, improves the robustness of the deep learning models. • A detailed comparison of deep learning models is conducted, considering the number of parameters, inference time, and efficiency, providing clear criteria for selecting the most suitable model for real-world applications.