Reliable Fruit Ripeness Classification from RGB Images via Calibrated Transfer Learning

View at publisher Summarise with VIRAT
Authors: Daniel Andrés Cevallos Valdiviezo, Erasmo García, Wilton Agila, Holger Cevallos
Cited by: 0
References: 11

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

We address nine-class fruit-ripeness recognition from commodity RGB images using a calibrated transfer-learning pipeline. EfficientNet-B0 is fine-tuned in two stages (with MobileNetV2 as a lightweight fallback) using class-weighted crossentropy with label smoothing and Adam with decoupled weight decay (AdamW); early stopping and stochastic weight averaging (SWA) promote generalization. At inference, mild test-time augmentation (TTA; horizontal flip and central crops) is combined with post-hoc temperature scaling (TS) to improve probabilistic reliability. On the Kaggle Fruit Ripeness dataset (apples, bananas, and oranges across {unripe, fresh/ripe, rotten}), the model achieved a Top-1 accuracy of 0.9735 with TTA (Top-3: 0.9995), micro-averaged receiver operating characteristic-area under the curve (ROC-AUC) of 1.000 (macro: 0.999), and reduced expected calibration error (ECE) from 0.044 to 0.004. Errors concentrated in visually adjacent maturity states within the same fruit (e.g., unripe vs. fresh/ripe), and Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations indicated that decisions relied on semantically meaningful regions. Implemented solely with standard TensorFlow/Keras components and commodity augmentations, the recipe provides a reproducible, deploymentoriented baseline that balances accuracy, efficiency, and calibrated confidence for food-quality applications in low-resource settings.

Fields of Study & Subjects

Smart Agriculture and AI Advanced Neural Network Applications Domain Adaptation and Few-Shot Learning

Publication details

Journal / source
International Conference on Artificial Intelligence in Information and Communication (ICAIIC)
Pages
257–262
Publication date
24 February 2026
Language
EN