Modul 39-Inf-VHM Vision in Human and Machine

Im Rahmen dieser Vorlesung lernen die Studierenden am Beispiel des visuellen Sytems des Menschen die äußerst fruchtbare interdisziplinäre Zusammenarbeit zwischen den experimentellen Neurowissenschaften und der technischen Bildverarbeitung mit neuronalen Architekturen kennen. Sie erwerben dabei Grundkenntnisse der Bildverarbeitung bezüglich Merkmalsextraktion, Objekterkennung und Segmentierung und lernen den aktuellen Stand der neurowissenschaftlichen Forschung zur biologischen Realisierung dieser Verarbeitungsprinzipien kennen. Praktische Übungen in MATLAB und Gruppenprojekte dienen der Vertiefung des erworbenen Wissen für Anwendungen aus der Computer Vision.

The target of this lecture is to introduce the fruitful interdisciplinary collaboration between neuroscience and computer vision for understanding the human visual system. The students acquire basic knowledge about feature extraction, object recognition, segmentation and learn about the current status of neuroscientific research on the biological realization of the underlying processing principles. Practical tutorials using MATLAB and group projects show the application of the learned concepts within the context of typical computer vision tasks.

This lecture gives an overview over the current state of knowledge about the human visual system and how this has led to new approaches to computer vision technology that are particularly suitable for embodied intelligent systems like humanoid robots. The lecture starts with an overview on the overall characteristics of human visual perception and a short review of the current state-of-the-art in computer vision. Then I will focus on the main human visual pathways for object recognition ("what") and spatial perception ("where") and present established models of early feature detection for these pathways. I will discuss the principle of redundancy reduction, which is an important concept for understanding sensory processing in the brain and explain methods like sparse coding for unsupervised learning of features. These methods have recently developed into well-established tools for general pattern recognition. Going from low-level perception to more high-level concepts, I will introduce the main models for object representation in the higher visual cortex and present corresponding hierarchical model implementations for object recognition which were shown to be very efficient in their application to humanoid robots. Another important topic will be the Gestalt laws of perception, and how the phenomena of perceptual grouping can be modeled using neurodynamical models for sensory segmentation. In the final part of the lecture I will focus on multi-modality and visual action-related representations like mirror-neurons in the brain and show how this has led to new learning and representation approaches for cognitive robots.

Grundkenntnisse in Mathematik: Mehrdimensionale Analysis

Basic required knowledge in mathematics: Multidimensional analysis

—