Correlation retinas measure the correlation product of an image projected on a sensor by optical means and a function f(x, y) stored in the retina. This optical correlation is well suited for the measurement of geometric moments, and this is why one can find in the literature several correlation retina circuits dedicated to it. Unfortunately, these architectures are not programmable and are dedicated to specific applications. Moment functions are of great interest in pattern recognition. This application needs to compute, for example, geometric moments whose order can depend on the application. Thus, the implementation of this function in a correlation retina requires a flexible architecture where the function f(x, y) can be modified to allow the measurement of the product of the image under analysis and f(x, y). The most robust solution is to memorize f(x, y) in memory devices distributed in the array of pixels constituting the retina. But, for technological problems, it is necessary to limit the number of bits used to store f(x, y) in the sensor. In this article, we propose to use dithering algorithms to code f(x, y) on only one bit. We present herein the architecture of the retina circuit on which we are currently working and show that it is possible to obtain approximate geometric moment values with it.