Tese de Doutoramento
A New Way of Performing Inspection Plans using Computer Vision and Deep Learning Algorithms
António Raimundo (Raimundo, A.);
Ano (publicação definitiva)
Mais Informação
Web of Science®

Esta publicação não está indexada na Web of Science®


Esta publicação não está indexada na Scopus

Google Scholar

Esta publicação não está indexada no Google Scholar

Industrial inspection is a crucial task to ensure the quality and safety of industrial processes. In recent years, deep learning models have shown promising results in performing these tasks. This thesis proposes a new deep learning architecture called YOLOX-Ray, specifically designed for industrial inspection. The YOLOX-Ray architecture is based on the family of You Only Look Once (YOLO) object detection algorithms and incorporates a new attention mechanism, SimAM, which allows for better feature extraction in the Feature Pyramid Network (FPN) and Path Aggregation Network (PAN). Additionally, a new cost function, Alpha-IoU, is implemented, which allows for better detection of objects on a smaller scale. The performance of the YOLOX-Ray architecture is evaluated in three case studies: hotspot detection, crack detection in infrastructure, and corrosion detection. Results show that the YOLOX-Ray architecture outperforms all other configurations, achieving mAP50 values of 89% for hotspot detection, 99.6% for crack detection, and 87.7% for corrosion detection. For the most challenging metric, mAP50:95, the values are 44.7%, 66.1%, and 51.8%, respectively. A comparative study analyzes the contribution of each component of the YOLOX-Ray architecture, showing that the combination of the SimAM attention mechanism and the Alpha-IoU cost function is essential to achieve the best performance. In summary, the proposed architecture’s ability to detect and locate multi-scale objects in industrial environments opens new possibilities for effective, efficient, and sustainable inspection processes in various industries, revolutionizing the field of industrial inspections.
Industrial Inspections,Computer Vision,Deep Learning,Object Detection,Attention Mechanisms
  • Engenharia Eletrotécnica, Eletrónica e Informática - Engenharia e Tecnologia

Com o objetivo de aumentar a investigação direcionada para o cumprimento dos Objetivos do Desenvolvimento Sustentável para 2030 das Nações Unidas, é disponibilizada no Ciência-IUL a possibilidade de associação, quando aplicável, dos artigos científicos aos Objetivos do Desenvolvimento Sustentável. Estes são os Objetivos do Desenvolvimento Sustentável identificados pelo(s) autor(es) para esta publicação. Para uma informação detalhada dos Objetivos do Desenvolvimento Sustentável, clique aqui.