Resumen:
This dissertation addresses the challenge of automating post-harvest fruit quality inspection
through the application of convolutional neural networks (CNNs) and deep learning models,
with a focus on ripeness, defects, and deformities. A systematic review establishes the state of
the art, identifying both advances and persistent gaps in CNN-based classification, including
dataset limitations, lack of alignment with international grading standards, and insufficient
robustness under real-world variability.
Building on this foundation, the study investigates banana ripeness classification under
varying illumination, demonstrating that although models achieve high accuracy on pristine
data, their generalization is significantly impaired under lighting changes. Gamma-based
augmentation improves robustness, yet model performance remains architecture-dependent,
underscoring structural limitations and the need for more resilient approaches.
Defect detection in apples and mangoes is examined through real and synthetic datasets,
leveraging RGB, silhouette, and spectral-band imagery. The introduction of a Multi-Input
architecture enhances discriminative power by integrating complementary representations,
achieving superior accuracy compared to traditional Single-Input configurations. Furthermore,
the exploration of spectral imaging highlights the potential of specific wavelengths for
improved defect visibility.
The study also introduces a comprehensive dataset for fruit deformity classification,
combining real, synthetic, and silhouette-based images of apples, mangoes, and strawberries.
Comparative evaluations of standard and lightweight CNN models reveal the effectiveness
of MobileNetV2, particularly in Multi-Input architectures, for capturing morphological
variations.
Overall, the contributions of this research include the development of robust datasets,
the proposal and evaluation of novel training strategies, and the validation of architectures
capable of addressing key challenges in fruit quality inspection. The findings provide
critical insights for advancing automated grading systems, highlighting trade-offs between
robustness, efficiency, and scalability, while laying the groundwork for future exploration of
transformer-based models, spectral imaging, and real-world deployment scenarios
