Computer Vision Methodology for Monitoring Morphological Degradation in Biopolymers Using EfficientNetB3 and U-Net Architectures

Cruz Aleman S.; Valdez Ponce D.; Carrizales Navarro J.; Figueroa Vilcarromero J.; Ronceros Rivas J. · International Conference on Electrical, Computer, Communications and Mechatronics Engineering, ICECCME 2025 (2025) · DOI: 10.1109/iceccme64568.2025.11277685

Resumen

This work introduces a dual-phase computer vision methodology for the automated monitoring of morphological degradation in biopolymer films. To address the need for accurate and scalable assessment, our approach integrates defect classification with semantic segmentation. In the first phase, we implement a classification model based on the EfficientNetB3 architecture using transfer learning. This model achieves a validation accuracy of 98.09%, outperforming reference architectures in similar tasks, such as DeepThin (93.5%), thereby setting a new state-of-the-art for crack detection on the reference dataset. The second phase focuses on precise damage quantification by implementing a U-Net architecture from its mathematical foundations. Trained on a dataset of 274 images, the U-Net model achieves an Intersection over Union (IoU) of 78.14%, validating its capability to accurately delineate degraded regions. This work confirms the feasibility of employing Convolutional Neural Networks (CNNs) for high-throughput defect detection and spatial quantification.

Instituciones: Universidad Peruana de Ciencias Aplicadas