Arma virumque cano, Trojae qui primus ab oris Italiamfato profugus, Laviniaque venit litora. This is the beginning of Ovid's story about Odysseus leaving Trojae to find his way home. I here tell about my own Odysee-like ex- riences that I have undergone when I attempted to simulate visual recognition. The Odyssee started with a structural description - tempt, then continued with region encoding with wave propagation and may possibly continue with a mixture of several shape descr- tion methods. Although my odyssey is still under its way I have made enough progress to convey the gist of my approach and to compare it to other vision systems. My driving intuition is that visual category representations need to be loose in order to be able to cope with the visual structural va- ability existent within categories and that these loose representations are somehow expressed as neural activity in the nervous system. I - gard such loose representations as the cause for experiencing visual illusions and the cause for many of those effects discovered in att- tional experiments. During my effort to find such loose represen- tions, I have made sometimes unexpected experiences that forced me to continuously rethink my approach and to abandon or turn over some of my initially strongly believed viewpoints.
The reader is presented an approach to the construction of a visual system, which is behaviorally, computationally and neurally motivated. The central goal is to characterize the process of visual categorization and to find a suitable representation format that can successfully deal with the structural variability existent within visual categories. It does not define such representations a priori but attempts to show directions on how to gradually work towards them. The book reviews past and existent theories of visual object and shape recognition in the fields of computer vision, neuroscience and psychology. The entire range of computations is discussed, as for example contour extraction in retinal circuits, orientation determination in cortical networks, position and scale independence of shape, as well as the issue of object and shape representation in a neural substrate. Region-based approaches are discussed and are modeled with wave-propagating networks. It is demonstrated how those networks operate on gray-scale images. A completely novel shape recognition architecture is proposed that can recognize simple shapes under various degraded conditions. It is discussed how such networks can be used for constructing basic-level object representations. It is envisioned how those networks can be implemented using the method of neuromorphic engineering, an analog electronic hardware substrate than can run neural computations in real-time and with little power.
