Welcome to the homepage of the lecture

Model-driven Deep Learning Lab for Image Analysis

Summer Term 2022

Model-driven Deep Learning Lab for Image Analysis

Lecturer: Karl Schrader

Examiner: Prof. Dr. Joachim Weickert
Summer Term 2022

Lecture (2h) with exercises (2h)
6 credit points

Online lecture with in-person tutorials
Lecture: Wednesday 12-14 c.t.
Tutorials: Monday 12-14 c.t., E1 3, room 016
First Lecture: Wednesday, April 13th, 2022



AnnouncementsDescriptionEntrance requirementsTutorialsExams
Lecture notes/Assignments



11/04/2022 The course is full and then some, as is the waiting list. While I greatly appreciate your desire to join the course, mailing me will not increase the amount of available slots.

06/04/2022 The course is full. Everybody who registered was added to the Team on MS Teams. If you registered before this date, please write a mail to Karl Schrader.

04/04/2022 Everybody who registered prior to 04/04/2022 17:00 was added to the Team on MS Teams. If you registered before this date, please write a mail to Karl Schrader.
Currently, there are 10 slots remaining.

04/03/2022 Website is online. Details about the teaching mode (in person/digital/hybrid) will be announced closer to the start of the semester.

Keep in mind that in most courses of studies, you also have to register via the HISPOS system of the Saarland University


Motivation: Neural networks take over an increasing number of tasks in image analysis, but their inner workings are a poorly understood black box for most architectures. Existing, model driven methods on the other hand are transparent in their modelling and offer robust convergence and stability theories, but designing performant and accurate numerical algorithms can require a lot of expertise. As a result, recent research attempts combine both worlds: Define the model as usual, but let a neural network take care of solving it.

Teaching Goals: The course is designed to give practical experience in designing, implementing and testing deep neural networks, but with applications far outside what you would find in a typical TensorFlow 101 course. It is designed to not require expensive hardware, making it as accessible as possible. After the lecture, participants should understand the basics of model-driven image analysis and be capable of building a neural network application for them from start to finish.

Contents: The lecture combines the necessary continuous and discrete theory to build simple denoising or inpainting models with the practical aspects of deep learning. It is seperated into four parts:

The first part of the lecture introduces the required basics of image analysis and introduces the models which we will need later on in the course. This includes fundamental aspects like the mathematical modelling of an image, the tasks of denoising and inpainting, as well as basic solvers for those models.

The second part introduces the required basics of neural networks. To build performant architectures later on, we need to discuss the fundamental building blocks and training approaches. Furthermore, we analyse the pros and cons of popular architectures and some connections to conventional solvers.

The third part concerns itself with the practical aspects of building and testing a neural network. We start from the very bottom: How can we get images from the disk into a neural network? How can we monitor a training run, and what can be concluded from those analytics? How can we store and later reuse a model?

The last part puts it all together: We design and implement deep neural network to solve the models we discusses in the first part. This requires an understanding of the model, the effects of different choices in the network design, and the ability to turn all that into code.


Basic mathematics courses (such as Mathematik für Informatiker I-III) are recommended. Understanding English is necessary. Image processing lectures such as "Image Processing and Computer Vision" are helpful for part 1, but not necessary. Machine learning knowledge or practical experience using TensorFlow can be helpful for parts 2 and 3, but are not required as well. For the programming assignments, some knowledge of Python 3 is required. You do not need a GPU or high performance PC.


There will be weekly programming assignments. During the weekly tutorial sessions, we discuss example solutions and further materials.

If you have questions concerning the tutorials, please do not hesitate to contact Karl Schrader.


Your grade will be determined by a final (solo) programming project at the end of the semester. You will be tasked to solve image analysis problems with methods from the lecture, along with a short report documenting your approach. The main project will run from July 13, 2022 till August 24, 2022. The second project will run from September 07, 2022 till October 19, 2022.
Please remember that you have to register online for the exam in the HISPOS system of the Saarland University at least one week before the start of the project period.

Lecture notes / Assignments

Lecture content will be available for download in MS Teams. This also holds for assignment submission and grading. You will be granted access to Teams after registration.


MIA Group
©2001-2020
The author is not
responsible for
the content of
external pages.

Imprint - Data protection