About the event
Inspection and characterization of materials are fundamental for understanding material properties, ensuring quality control, and accelerating materials development and advanced manufacturing. However, conventional approaches often require extensive expert analysis of complex images, making them time-consuming and difficult to scale. With the rapid growth of high-resolution imaging techniques, deep learning-based methods provide powerful solutions for automated materials image analysis by learning meaningful visual features for tasks such as defect classification, microstructure recognition, process monitoring, and intelligent quality assessment.
This workshop offers a focused, hands-on introduction to deep learning for visual inspection in materials science. Using the Severstal Steel Defect Detection dataset as a concrete, running example, participants will follow a complete analysis pipeline, from raw image data through model training to interpretable results. The techniques covered are deliberately chosen for their breadth of applicability: while the examples are grounded in industrial quality control, the same approaches translate directly to defect detection in electron microscopy, anomaly identification in medical imaging, and structural characterization across a wide range of domains.
Over three half-day sessions, participants will build solid intuition for how modern neural networks process and learn from image data, implement deep neural network classifiers from scratch, apply transfer learning to leverage pre-trained models for materials-specific tasks, and explore state-of-the-art architectures (including convolutional networks and transformer-based approaches) used across both research and industry.
We encourage participants to follow up in our upcoming workshop session on Sept. 30th to dive deeper into object detection and instance segmentation.
Who is this course for?
-
Engineers, analysts, academic researchers, and students in materials science or manufacturing who want to apply AI-based image analysis to quality control and defect inspection
-
Anyone with a general interest in computer vision and deep learning, regardless of domain
Key takeaways for participants
-
Understand the core concepts behind deep neural networks (including CNNs) and transfer learning and know when and why to use them for image-based inspection tasks
-
Gain hands-on experience building and training defect classifiers — from a simple binary classifier to multi-class CNNs
-
Learn how to extract and visualize feature embeddings to understand what your model has learned
-
Know how to leverage pretrained models (VGG, ResNet, ViT) through transfer learning, even with limited amount of data
Prerequisites
-
Basic Python programming (loops, functions, libraries)
-
Familiarity with NumPy or similar data manipulation tools is helpful but not required
-
No prior deep learning experience necessary – core concepts will be introduced from scratch
Schedule
| Day | Time | Contents |
|---|---|---|
| 22 September 2026 | 09:00 – 12:00 | Anomaly Detection & Feature Extraction Introduction to industrial inspection and defect classification. Build a binary defect classifier, explore learned embeddings, and visualize feature clusters |
| 23 September 2026 | 09:00 – 12:00 | CNN Fundamentals & Multi-class Classification Dive into convolutional neural networks — convolutions, pooling, and the LeNet architecture. Apply training best practices |
| 24 September 2026 | 09:00 – 12:00 | Transfer Learning & Advanced Vision Architectures Explore pretrained deep networks (VGG, ResNet) and Vision Transformers (ViT). Fine-tune models on defect data and compare performance across architectures. |
Instructors and Teaching Assistants
-
Benedikt Neyses (Mimer AIF / RISE)
-
Andreas Thore (Mimer AIF / RISE)
-
Marzieh Saeedimasine (MIMER AIF/NAISS)
-
Ruiwen Xie (MIMER AIF/NAISS)
-
Smita Chakraborty (Mimer AIF / RISE)
-
Yuvarajendra Anjaneya Reddy (Mimer AIF / RISE)
-
Yonglei Wang (MIMER AIF/NAISS)