Programmable optical CNN implementation based on the template pixels' angular coding

Abstract

Within the programmable opto-electronic analogic computer (POAC) framework a new, feed forward only optical CNN-UM implementation has been introduced. It is grounded on an innovative semi-incoherent optical correlator architecture. Angular coding of the template pixels determines the operation of this optical CNN implementation, therefore it is real time and flexibly programmable. We have demonstrated its feasibility and operation by an experimental setup. Our correlator architecture makes it possible to execute algorithms real time, which cannot be done by any other existing optical correlator so far. Our architecture unifies the advantages of coherent and incoherent optical correlators, provides a more robust frame and avoids their main hindrances. In the POAC framework the resulting correlogram is measured by a programmable adaptive sensor array, a special visual CNN-UM chip. So, local parallel programs fulfill both the necessary pre and post processing with the required adaptive thresholding. However, because of the limited resolution of available visual CNN chips (28/spl times/28), all-optical optical pre- and post-processing will be used, as well.