Future Work

The most interesting piece of future work would be to implement the network on a physical integrated circuit. There are many parameters that SPICE cannot take into account that affect the performance of an actual silicon die. I have not designed a physical chip layout, which would be important in order to assess the feasibility of implementing this network for large arrays in high-definition applications.

The anticipated improved motion detection of the new network has not been tested. It would be worth simulating the effect of exposing the network to a changing input. It is also important to quantify the response time of the network to changing input. While it is unlikely that this kind of system would need to deal with images changing quicker than about 50 frames per second, which should be easily achievable, this has not been quantified. Although the individual circuit elements are simple, signals from each pixel have to propagate across the entire network, and so the settling time can be expected to be of the order of microseconds at least.

In this project I assumed that the input pixel voltages were linearly related to intensity. However, in some systems the voltages are logarithmically related [2]. This allows the system to operate over a wide range of input intensities, for example operation during both the day and night. This is also the strategy used by the human retina [13]. An examination of the performance of the system under these conditions would be valuable.

More algorithms need to be developed that perform higher level functions. The greatest potential for this network is as a preprocessor to further processing. There are already systems that detect optical flow, for example [14]. These systems detect motion and give an output signal that is a vector representing the direction in which motion is taking place. This system could be combined with such a network. Edge detection provides valuable information for detecting motion because edges are the easiest features to detect, while motion is one of the best ways to detect edges. Motion is a good way to detect the boundaries of strongly textured or camouflaged regions. Such regions can usually only be seen by humans when the regions move with respect to each other, as when a well camouflaged animal moves. Thus there is the potential for two systems to co-operate and improve the performance of each other.

Matthew Exon 2004-05-23