Time-derivative cellular neural networks and their applications

Biological organisms excel at processing large amount of data in real time whereas computers usually are not good at such tasks because of their point processor architecture. Thus a great deal of effort has been spent on the development of machines inspired by biological structures in computer vision studies. Such vision machines require the simple low-level feature detectors similar in that of the retina. The low-level feature detectors measure local image properties as scale, orientation, and velocity. It has been shown that active resistive diffusion networks with low connectivity offer a common framework for the implementation of the low-level feature detectors commonly used in vision. Cellular Neural Networks (CNNs) consist of regular arrays of simple processing units that interact with only their nearest neighbors. The CNN architecture bears striking resemblance to aforementioned biological organisms and they are tailor-made for analog VLSI implementations because of their nearest neighbor connections and usually space invariant connection weights. The scope of this thesis is time-derivative CNN (TDCNN) architecture which is the general case of CNN and has been introduced recently in literature. Firstly, the simulation method of TDCNN which combines forward and backward Euler approximations is examined. The empirical condition which must be met for the method is proved. In addition, the TDCNN structure, which has only been used for the realization of 3-D band-pass filter in literature so far, is employed in the realization of a 3-D spatio-temporal Gabor-type filter. The TDCNN realization of 3-D spatio-temporal Gabor-type filter is simpler and more straightforward than that of the 3-D band-pass filter, since there are many more variables using in the 3-D band-pass filter realization and also their function and the process of tuning the filter are not clear. What is more, it is shown that the TDCNN realization of the other 3-D spatio-temporal filters could be easily done with the existing CNN templates in literature by using the method which is used for TDCNN realization of 3-D spatio-temporal Gabor-type filter.