DEEP Inspection for Materials Science
Learn about visual inspection, object detection, and instance segmentation in materials science using deep learning methods.
Inspection and characterization of materials, and detecting defects in materials are fundamental for understanding material properties, ensuring quality control, and accelerating materials development and advanced manufacturing. Instead of conventional approaches requiring extensive expert analysis of complex images, deep learning-based methods provide powerful solutions for automated materials image analysis. This workshop, organized by the Sweden AI Factory, offers a focused, hands-on introduction to deep learning for visual inspection, object detection, and instance segmentation in materials science.
Key takeaways
By the end of this course, you will be able to:
- Explain the two core assumptions behind CNNs (translation invariance and locality) and why they make convolution more parameter-efficient than fully connected layers
- Describe and distinguish four computer vision paradigms: anomaly detection, image classification, object detection, and image segmentation, and identify which is appropriate for a given inspection task
- Apply transfer learning (feature extraction and fine-tuning) with pretrained CNN and transformer architectures (VGG, ResNet, ViT) to a new defect classification dataset
- Explain how PatchCore detects anomalies without labeled defect examples, and contrast this with supervised classifiers and their open-set failure modes
- Understand the YOLO family’s progression from two-stage pipelines to single-pass anchor-free detection, and apply YOLO11 to both object detection and instance segmentation tasks
- Distinguish semantic segmentation (U-Net, SegFormer) from instance segmentation (YOLO11-seg) and select the appropriate architecture for a given materials inspection scenario
Who is this for?
- Practitioners and researchers in materials science and industrial inspection
- Data scientists working on deep learning-based applications
- Engineers, analysts, and domain specialists who need to move beyond pass/fail classification toward precise spatial localization of features within images
Prerequisites
- Basic Python programming skills and familiarity with packages like NumPy, Pandas, and Matplotlib
- Experience working with Jupyter notebooks (recommended but not mandatory)
- No prior deep learning experience necessary – core concepts will be introduced from scratch