Identification of lines is a basic machine vision problem which is essential in a large number of applications. The Hough transform is ideal for line detection because it is robust, relatively insensitive to noise and degrades gracefully to occlusions in addition to other advantages. But it has the disadvantage of being computationally intensive, so it is relatively slow, and for this reason it has found little use in real time applications where its unique abilities would allow it to have wide utility. This handicap could be removed if practical implementation with artificial neural networks were possible. Example system applications include autonomous navigation, tracking multiple targets, curve following, mensuration and image recognition.
Neural-like analog circuitry is suggested for the image to parameter space mapping while a Hopfield-type network is proposed for the parameter space peak detection. A relatively simple addition to the basic system allows for multiple peak detection. Practical circuit details are discussed and a new circuit model for the Hopfield neuron is presented. Network performance under worst case component tolerances and fault conditions is examined. A hardware/software approach is presented to improve the fault tolerant characteristics of the network. A twenty-five pixel implementation was constructed to verify simulation and theoretical results. Solution time under thirty microseconds is offered with general purpose operational amplifiers.