Methodology

A resistive fuse grid is a fairly simple circuit that effectively defines any large intensity difference between two adjacent pixels as an object boundary. These boundaries define objects. Within each object, a simple smoothing process is used to eliminate noise. This approach is simple and useful, but there are common situations which it does not handle adequately. In particular, an isolated pixel which has a totally incorrect intensity value, maybe because its receptor is broken, or because of dust on the lens, will always be classified as an object in itself, whereas it really should be returned to a similar value to its neighbours. Furthermore, noisy pixels that are close together have a tendency to reinforce each other.

The circuit described in this project will use more contextual information to detect and correct such situations. In current networks, each resistive fuse effectively only uses information from two pixels to make its segmentation decision. By using information from pixels further away, it should be possible to eliminate some or all of the problems outlined above.

The circuit that was designed was implemented only as a SPICE simulation, not as a physical circuit. The reason for this is that the time available could be more fruitfully employed optimising and analysing the design. The circuit is so complex that actually constructing it by hand would take an inordinate amount of time. Furthermore, such a circuit could only be built on a fairly small scale, of the order of 8 pixels square, which would only allow the most rudimentary testing. Implementation of the circuit in VLSI, as intended by the design, is too expensive for the limited scope of this project. Implementation of the circuit would be most useful as part of a larger integrated circuit with other processing elements.