http://www.dspace.espol.edu.ec/handle/123456789/69944 FIEC 2026-09-09T10:03:15Z http://www.dspace.espol.edu.ec/handle/123456789/69947 Title: External Quality Inspection of Post harvest fruits using Computational Intelligence Methods Authors: Chuquimarca Jiménez, Luis Enrique; Vintimilla Burgos, Boris Xavier, Director Abstract: 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 2026-01-01T00:00:00Z