Development of computer vision based models for automated crack detection

• Ano de defesa: 2020
• Autor(a) principal: CUNHA, Beatriz Sales da
• Orientador(a): MOURA, Márcio José das Chagas
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de Pernambuco
• Programa de Pós-Graduação: Programa de Pos Graduacao em Engenharia de Producao
• País: Brasil
• Palavras-chave em Português: Engenharia de Produção; Detecção de trincas; Processamento de imagens; Segmentação; Análise de textura; Aprendizagem de máquina
• Link de acesso: https://repositorio.ufpe.br/handle/123456789/37961

Systems subjected to continuous operation in harsh conditions are exposed to different failure mechanisms (e.g., corrosion, fatigue, and temperature-related defects). In this context, inspection and health monitoring have become crucial to prevent system, environment, and users from severe damage. However, visual inspection strongly depends on human’ experience, having its accuracy influenced by the physical and cognitive state of the inspector (i.e., human factors). Particularly, infrastructures need to be periodically inspected, which is costly, time-consuming, hazardous and biased. Nowadays, the increase in computer power allows for analyzing a considerable number of images in a shorter time and use more robust algorithms. Advances in Computer Vision (CV) and Machine Learning (ML) provide the means to the development of automated, accurate, non-contact and non-destructive inspection methods. Therefore, this dissertation proposes and compares the adoption of different CV approaches to extract features for crack detection. In fact, we applied texture-based and region-based methods to a real concrete crack image database, and then the results fed four ML models to identify crack existence, namely Support Vector Machine (SVM), Multilayer Perceptron (MLP), Adaboost (AB), and Random Forest (RF). Results show the potential of data preprocessing to improve methods’ performance in reaching a balanced accuracy above 97%.