Ensemble de Redes Convolucionais para Inspeção Visual de Estruturas de Concreto Armado

• Ano de defesa: 2025
• Autor(a) principal: DINIZ, Joel de Conceição Nogueira
• Orientador(a): PAIVA, Anselmo Cardoso de
• Banca de defesa: PAIVA, Anselmo Cardoso de , CUNHA, Sandra Cristina Alves Pereira da Silva , BRAZ JUNIOR, Geraldo , ALMEIDA, João Dallyson Sousa de , CONCI, Aura
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Maranhão
• Programa de Pós-Graduação: PROGRAMA DE PÓS-GRADUAÇÃO DOUTORADO EM CIÊNCIA DA COMPUTAÇÃO
• Departamento: DEPARTAMENTO DE INFORMÁTICA/CCET
• País: Brasil
• Palavras-chave em Português: Patologias no Concreto; Aprendizado Profundo; Visão Computacional; Detecção; Classificação
• Palavras-chave em Inglês: Deep Learning; Computer Vision; Detection; Classification; Concrete Pathologies
• Área do conhecimento CNPq: Estruturas de Concreto
• Link de acesso: https://tedebc.ufma.br/jspui/handle/tede/6651

Pathologies in concrete structures can be visually detected on their surface, such as cracks or fissures, fragmentation of concrete, efflorescence, corrosion stains, and exposed steel bars (the latter two occurring in reinforced concrete). Therefore, these pathologies can be analyzed via images of reinforced concrete structures. This thesis proposes an ensemble of convolutional networks for visual inspection of reinforced concrete structures. This method speeds up the detection task and increases its effectiveness by saving time on the identification to be analyzed and by eliminating or reducing errors, such as those arising from human error during the massive execution of tedious tasks in the analysis. The task of identifying pathology can be performed using a convolutional neural network if the images are cropped to the specific pathology to be identified, or using a detection network if the images are broad and the pathology is inserted into a context with several classes of pathology, including areas without pathology. Another task that enables the identification and analysis of these pathologies is segmentation. The method was tested with detection and classification tasks. The neural network architectures used for detection were YOLO v11 and TOOD (Taskaligned One-stage Object Detection) for the single-stage approach, and Faster RCNN for the two-stage approach. The three networks were then combined for the Ensemble application, using Weighted Box Fusion. For the classification task, the network architectures used were DenseNet121, ResNet50, and MobileNetV3. Research was conducted to locate and select appropriate datasets for the proposed method. The dataset selected for classification is Ozgnel, and the one for detection is CODEBRIM. The detection task allows an artifact to be located and classified. Although this approach is efficient, the research tested using detection to locate pathologies without initially defining which class they belong to, and then using a specific classification neural network to define the types of pathologies and, above all, to eliminate false positives. The approach of combining a detection network and a dedicated classification network achieved the desired result, with Ensemble used to increase sensitivity in the detection phase, thereby increasing the number of artifacts located. This approach allows false positives to pass in the first phase, but they are eliminated in the last phase.