We address the automation of visual inspection of machined metal products using machine vision, with the aim of increasing speed and objectivity. The focus is on porosity
detection, where surface reflectivity and curvature pose a challenge. An experimental setup is developed to ensure repeatable image acquisition of adequate quality. Images
of real samples are prepared, annotated, and used to compare different machine vision methods. Classical approaches, a convolutional neural network, and two unsupervised anomaly detection methods are compared, with emphasis on the latter. Classification performance is evaluated, while the influence of key factors, such as acquisition conditions, surface contaminants, and model parameter settings, is analysed. The results demonstrate the potential of the proposed approach, with Patch Distribution Modeling (PaDiM) identified as the most appropriate method, achieving an F1 score of 0.96.
|
