|
“Omnes enim causae effectum naturalium habent dari per lineas, angulus et figuras. Aliter enim impossibile est scire «propter quid» in illis.”
Robert Grosseteste (1168 - 1253)
My research concerns a variety of topics related to the computational analysis and interpretation of visual images. Such research is inherently interdisciplinary, including mostly areas of computer vision and image processing, but often overlapping with visual perception and cognition, pattern recognition, and physics of light reflection.
It is my firm belief that solutions implemented in any information-processing device, supposed to simulate corresponding human functions, will eventually converge to mimic the workings of the human brain (in terms of the approach to analysis and the organizational structure, and thus efficiency, distribution of resources, etc.). From that standpoint, the process of constant (human) optimization of such solutions resembles the one of biological evolution, which led to the development of the human brain1.
Among the plethora of processes performed by the brain, my work concentrates on modeling vision, the youngest and most complex of the senses, dominating our perception and constantly requiring some 40% of all brain capacity. In a similar vein as mentioned above, I believe that the development of computer algorithms, aimed at interpreting visual scenes, should closely follow the processing in the human visual system. This is especially true if results of those algorithms are to be presented to a human user. Nevertheless, human vision has been optimized to interpret the input with more-or-less a single evolutionary goal: to make the best possible movement in a given situation. Therefore, it is sensible to closely follow the brain's model when the program's functionality should resemble that of a human. Otherwise, better solutions might be possible and could be sought.
Three primary goals provide the motivation for my work:
1) personal fascination and enthusiasm for acquiring knowledge
From what is known about the human visual system, it is clear that deciphering it has to be a long-term goal, unlikely to be achieved within our life-span. Nevertheless, even if numerous attempts to make ground-breaking contributions to the field come to dead ends, the learning experience and the acquired knowledge provide sufficient satisfaction and motivation for further research. The fact that existing computer vision methods might still be very primitive compared to the abilities of human vision is consequently viewed as a challenge (to reach a seemingly unreachable goal), and not as a discouragement.
2) construction of models for the representation of visual information
If understanding and modeling of the human brain is our ultimate goal, we would like to have a sound theoretical framework inside which visual processes could be explained. This theoretical basis would enable us to analyze, and possibly predict, the behavior of the human visual system. Moreover, it would facilitate the transfer of knowledge from biological to computer vision systems, and indicate the correct approaches to solving particular problems.
3) creation of novel algorithms for applications in image understanding
Nurtured by the acquired knowledge and understanding of human vision, we would ideally want to apply the same principles in computer vision. Analysis of the principles of human visual system should provide us with two equally important insights in this direction. The first is about the necessary conditions. In other words, the first goal of the computer vision system is to offer functionality which is considered elementary by the human viewers. The second insight that we hope to acquire is about the sufficient conditions. Put simply, the second important goal of any computer vision system is to provide only what is within the perceptual realm of the human observer, exploiting the limitations of the human visual system whenever possible in order to simplify or speed-up visual processing.
With these motivating factors, I lack neither the fascination for my field nor the drive to engage in further scientific research.
1
A large number of compelling examples of the correspondence between neural processing on one, and computational methods derived by humans on the other side, is presented in the book "Vision Science - Photons to Phenomenology" (1999) by Stephen E. Palmer, as well as works by Hubel&Wiesel, David Field, Eero Simoncelli, David Mumford, David Lowe, the scale-space proponents Andrew Witkin, Jan Koenderink and Tony Lindeberg, and many others.
|
|