Reverse Engineering the Principal Image Processing Architectures of the Macula Lutea Within the Human Retina

摘要

We present here a biomorphic CMOS colour opponent retinal processing algorithm and chip, representing the image-processing of the human Macula Lutea, with simulated and measured results. This chip has colour selective photodiodes (representing blue and red retinal cone cells) implemented without any post processing, using the intrinsic absorption of silicon as colour filter, and allowing double colour opponent receptive field implementation. Utilising two convolution stages (an improved resistive network with feedback that inhibits lateral spreading under high intensity light conditions, and a current-mode bidirectional 3×3 distributed reduced Laplacian filter), allowing asymmetric and effective Laplacian filter implementations of any size from 3×3 to larger than the array itself. The current-mode circuitry represents the Macula ganglion, bipolar cell interface, and the resistive network high light intensity inhibition has been observed within retinal horizontal cell networks. This work is directly relevant to distributed focal plane image processing systems, either as standalone feature extraction devices where low space and power are essential, or as a retinal replacement aid for the visually impaired.