Hardware Implementation of Geometric Moment Functions in a CMOS Retina

摘要

We present in this paper a new method for implementing geometric moment functions in a CMOS retina. It is based on the computation of the correlation value between the image under analysis and a second image since there is a similarity between the expression of the moment of an image and that of the correlation of two images. The second image which is stored in memory devices in the circuit is approximated by a binary image using a dithering algorithm in order to reduce hardware implementation cost. As a result the value of the moment is also an approximate one. Computer simulations using the COIL 100 Columbia image database on 128x128 pixel images show that the maximal relative error between the approximate and the exact value is less than 1% for moments of order less than 2, and less than 5% for moments of order less than 6. Finally, we have considered an object localization application and quantified the error in the localization due to the use of the approximate moment values instead of the exact values.