Seminar by Abdel Rahman Dakak

Deep Learning-Based Defect Detection in CT Volumes of Aluminium Alloy Castings

Deep learning (DL) has become one of the most widely used artificial intelligence techniques in the field of automation. Today, DL models can interpret images in the same way our brains do, such as face recognition, anomaly detection or self-driving cars, as these tasks are too difficult to solve with traditional algorithms in terms of processing, execution time and implementation settings. Training DL models can be defined as the process by which an artificial neural network learns from a set of examples to perform a specific task. This seminar presents a brief introduction to deep learning applied to images and a case study on industrial X-ray computed tomography (CT) data of metal castings.
Since CT volumes are highly prone to artifacts that can be mistaken for defects by conventional image processing algorithms, a deep learning approach was developed to analyze defects in these volumes based on a three-step pipeline: (1) 2D segmentation of CT slices with deep segmentation to detect suspicious discontinuities; (2) classification of these discontinuities into defects or artifacts using a trained neural network classifier; (3) localization of the validated defects in 3D to classify them as porosities or shrinkage cavities using a siamese neural network.

Keywords: Machine Learning, Deep Learning, Casting, CT, X-Ray, Defect Detection, Few-Shot Learning.