Home | Impressum | Datenschutz | KIT

Machine Vision

Machine Vision
Typ: Vorlesung Links:
Semester: Wintersemester
Ort:

Do, 11.40 Johann-Gottfried-Tulla-Hörsaal

Fr, 30.22 Gaede-Hörsaal

Zeit:

Do, 11:30 - 13:00 Uhr und

Fr, 14:00 - 15:30 Uhr

Beginn: 18.10.2018
Dozent:

Dr. rer. nat Martin Lauer
M.Sc. Jannik Quehl

SWS: 4 h
LVNr.: 2137308

Prüfung / Exam

Die Klausur zur Vorlesung "Machine Vision" (Dr. Lauer) findet am 07.08.2018 in der Zeit von 08:30 bis 09:30 Uhr statt. Weitere Informationen zur Klausur finden Sie hier.

The written exam of the lecture "Machine Vision" (Dr. Lauer) will be held on 07th of August, 2018 in the time between 08:30 and 09:30. Futher information about the exam can be found here.

Content

Machine vision (or computer vision) describes all kind of techniques that can be used to extract information from camera images in an automated way. Considerable improvements of machine vision techniques throughout recent years, e.g. by the advent of deep learning, have caused growing interest in these techniques and enabled applications in various domains, e.g. robotics, autonomous driving, gaming, production control, visual inspection, medicine, surveillance systems, and augmented reality.

The lecture on machine vision covers basic techniques of machine vision. It focuses on the following topics:

  • image preprocessing
  • edge and corner detection
  • curve and parameter fitting
  • color processing
  • image segmentation
  • camera optics
  • pattern recognition
  • deep learning

Image preprocessing:
The chapter on image processing discusses techniques and algorithms to filter and enhance the image quality. Starting from an analysis of the typical phenomena of digital camera based image capturing the lecture introduces the Fourier transform and the Shannon-Nyquist sampling theorem. Furthermore, it introduces gray level histogram based techniques including high dynamic range imaging. The disussion of image convolution and typical filters for image enhancement concludes the chapter.

Edge and corner detection:
Gray level edges and gray level corners play an important role in machine vision since gray level edges often reveal valueable information about the boundaries and shape of objects. Gray level corners can be used as feature points since they can be identified easily in other images. This chapter introduces filters and algorithms to reveal gray level edges and gray level corners like the Canny edge detector and the Harris corner detector.

Curve and parameter fitting:
In order to describe an image by means of geometric primitives (e.g. lines, circles, ellipses) instead of just pixels robust curve and parameter fitting algorithms are necessary. The lecture introduces and discusses the Hough transform, total least sum of squares parameter fitting as well as robust alternatives (M-estimators, least trimmed sum of squares, RANSAC)

Color processing:
The short chapter on color processing discusses the role of color information in machine vision and introduces various models for color understanding and color representation. It concludes with the topic of color consistency.

Image Segmentation:
Image segmentation belongs to the core techniques of machine vision. The goal of image segmentation is to subdivide the image into several areas. Each area shares common properties, i.e. similar color, similar hatching, or similar semantic interpretation. Various ideas for image segmentation exist which can be used to create more or less complex algorithms. The lecture introduces the most important approaches ranging from the simpler algorithms like region growing, connected components labeling, and morphological operations up to highly flexible and powerful methods like level set approaches and random fields.

Camera optics:
The content of an image is related by the optics of the camera to the 3-dimensional world. In this chapter the lecture introduces optical models that describe the relationship between the world and the image including the pinhole camera model, the thin lens model, telecentric cameras, and catadioptric sensors. Furthermore, the lecture introduces camera calibration methods that can be used to determine the optical mapping of a real camera.

Pattern recognition:
Pattern recognition aims at recognizing semantic information in an image, i.e. not just analyzing gray values or colors of pixels but revealing which kind of object is shown by the pixels. This task goes beyond classical measurement theory and enters the large field of artificial intelligence. Rather than just being developped and optimized by a programmer, the algorithms are adapting themselves to their specific task using training algorithms that are based on large collections of sample images.

The chapter of pattern recognition introduces standard techniques of pattern recognition in the context of image understanding like the support vector machine (SVM), decision trees, ensemble and boosting techniques. It combines those classifiers with powerful feature representation techniques like the histogram of oriented gradients (HOG) features, locally binary patterns (LBP), and Haar features.

Deep learning:
Throughout recent years standard pattern recognition technqiues have more and more been outperformed by deep learning techniques. Deep learning is based on artificial neural networks, a very generic and powerful form of a classifier. The lecture introduces multi layer perceptrons as the most relevant form of artificial neural networks, discusses training algorithms and strategies to achieve powerful classifiers based on deep learning including deep auto encoders, convolutional networks, and multi task learning, among others.


The lecture is composed out of some theoretical parts and more practical parts. While the theoretical  parts introduce and discuss new algorithms the practical parts of the lecture show example implementations and allow the students to get hands-on experience on machine learning techniques.