The role of local scale and orientation in feature location using neural nets

The application of one pass orientation and scale selective filters to feature location is investigated. The particular case study developed deals with the location of eyes in head-and-shoulders images using artificial neural networks trained by back-error-propagation. Three types of filters were studied. Conventional Marr edge detectors, Gabor filters selective to the horizontal and vertical directions, and also edge detectors which extract high resolution information along each of the two directions using 2D separable wavelet filters. Tests were conducted to locate the centre of the pupil in the right eye in sixty head-and-shoulders images. The results are compared for the different filters, and also using the raw pixel image directly.<>