Thesis Tratamiento de sombras para detectar deterioros en pavimentos con machine learning
Loading...
Date
2025-08
Journal Title
Journal ISSN
Volume Title
Program
Ingeniería Civil Informática
Departament
Campus
Campus Santiago San Joaquín
Abstract
En una metodología para la evaluación de deterioros en imágenes de alta resolución de pavimentos urbanos se busca corregir errores de detección causados por la presencia de sombras, para ello se compararon seis métodos de eliminación de sombras aplicados a tres datasets representativos de distintos pasos de la etapa de preprocesamiento. A continuación, las imágenes resultantes se procesaron con YOLOv5, el algoritmo detector de deterioros, para evaluar si hay mejoras en su precisión. De los dos métodos de procesamiento de imágenes y los cuatro basados en Deep Learning, se comprueba que DC-ShadowNet, un modelo basado en aprendizaje no supervisado, al aplicarse en las imágenes de pavimentos antes de la etapa de normalización, evitó que YOLOv5 confundiera las sombras con deterioros.
A methodology for assessing deterioration in high-resolution images of urban pavements seeks to correct detection errors caused by the presence of shadows. To this end, six shadow removal methods were compared, applied to three representative datasets from different steps of the preprocessing stage. The resulting images were then processed with YOLOv5, the deterioration detection algorithm, to assess whether there were improvements in accuracy. Of the two image processing methods and the four based on Deep Learning, it was found that DC-ShadowNet, a model based on unsupervised learning, when applied to pavement images before the normalization stage, prevented YOLOv5 from mistaking shadows for deterioration.
A methodology for assessing deterioration in high-resolution images of urban pavements seeks to correct detection errors caused by the presence of shadows. To this end, six shadow removal methods were compared, applied to three representative datasets from different steps of the preprocessing stage. The resulting images were then processed with YOLOv5, the deterioration detection algorithm, to assess whether there were improvements in accuracy. Of the two image processing methods and the four based on Deep Learning, it was found that DC-ShadowNet, a model based on unsupervised learning, when applied to pavement images before the normalization stage, prevented YOLOv5 from mistaking shadows for deterioration.
Description
Keywords
Eliminación de sombras, Pavimentos urbanos, Imágenes 2D de alta resolución, Procesamiento de imágenes, Deep learning, Shadow removal, Urban pavements, High-resolution 2D images, Image processing