Banca de DEFESA: Vinícius Araújo Gonçalves

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : Vinícius Araújo Gonçalves
DATE: 22/05/2025
TIME: 16:00
LOCAL: Videoconferência
TITLE:

Detection of Window Frames in Public Buildings Using Image-Based Deep Learning Techniques

KEY WORDS:

Neural Network; Machine Learning; Deep Learning; Convolutional Network.

PAGES: 121
BIG AREA: Ciências Sociais Aplicadas
AREA: Arquitetura e Urbanismo
SUMMARY:

The use of image processing techniques (IPTs) has proven to be promising in identifying construction-related issues, with the potential to reduce the need for in-person inspections conducted by specialists. These IPTs are primarily employed to manipulate images and extract features related to defects and pathological manifestations in buildings. However, the widespread adoption of these techniques faces challenges due to variations in real-world conditions, such as changes in lighting and the presence of shadows. To overcome these limitations, this research proposes a method based on computer vision, utilizing a deep convolutional neural network (CNN) architecture to detect window frames in public buildings. The deep learning model was trained using a specific dataset, consisting of images of public buildings from all 26 Brazilian states and the Federal District, totaling over 1,840 constructions. The CNN was trained with more than 19,000 images at a resolution of 227 × 227 pixels, achieving an accuracy of approximately 80%. The robustness and adaptability of the method were tested on high-resolution images of varying sizes, captured under different lighting conditions and featuring various types of window frames that were not used in the model’s training and validation phases. In these scenarios, the model reached an accuracy of 83.34%. The results demonstrate the effectiveness of the proposed approach, highlighting its potential for the automatic detection of window frames and for generating quantitative surveys aimed at their restoration. The presented technology stands out for its satisfactory performance and its applicability in real-world scenarios.

COMMITTEE MEMBERS:
Presidente - 2352146 - JOAO DA COSTA PANTOJA
Interno - 2212892 - MARCIO AUGUSTO ROMA BUZAR
Externo à Instituição - MARCO AURELIO SOUZA BESSA - GDF
Externa à Instituição - NATHALY SARASTY NARVAEZ - UnB