A silicon retina is a VLSI (very large scale integration) chip that has been designed based on the neural architecture and functionality of a biological retina. Such devices are expected to be useful as a novel vision sensors for autonomous mobile robots due to their power efficiency, compactness, and real-time processing ability. In addition, their low noise output and wide operating range are both extremely important especially for practical applications.

The logarithmic response of the silicon retina helps it to exhibit brightness constancy. The step chart is illuminated using fluorescent lamp, and its intensity is varied from 750lx to 1500lx. The top and bottom rows of images are outputs of silicon retinas with linear and logarithmic responses, respectively. The logarithmic response is almost constant against the variation in illumination.

Kameda et al. developed a frame-based silicon retina with sustained and transient responses.¹ This chip uses an active pixel sensor (APS): a conventional, sampled photosensor consisting of a photodiode and a source-follower circuit. The APS is highly light sensitive because photoelectrons are accumulated in the parasitic capacitor of the photodiode. In addition, the outputs of the chip are offset-suppressed analog voltages, with embedded sample/hold circuits to eliminate both the pattern noise due to the amplifier offset and the fixed-pattern noise of the APS. In this chip, the output voltage of the APS is linearly proportional to the input light intensity, and its sensitivity can be controlled by changing the accumulation time. However, the dynamic range for individual images is limited.

We recently fabricated a wide-dynamic-range silicon retina based on the circuit design of this chip. In order to widen the dynamic range, however, we used the stepped-reset-gate-voltage technique,² where the voltage signal fed to the gate node of the reset transistor in the APS is changed during the accumulation period. One of the advantages of this technique is that the response characteristics can be controlled by an external signal. Using this technique, we approximated a logarithmic APS response.

More importantly, the new chip improves both the dynamic range and robustness against changes in illumination. By combining the logarithmic response of the photosensor and Laplacian-Gaussian-like spatial filtering, brightness constancy is produced. The principle behind this is well known—Land's retinex theory³—and its implementation using resistive networks has been reported by Moore et al.⁴ Figure 1 shows the brightness constancy of the new chip. In our experiment, a density step chart placed in front of the silicon retina camera was illuminated by a fluorescent lamp with a dimmer control. We measured the response of the silicon retina while varying the light intensity of the fluorescent lamp from 750 to 1500 lx. The output of the device with the logarithmic response is almost constant, while with the linear response it changes significantly and is saturated at high light intensities. Further, the transient response of the chip exhibits similar brightness constancy.

Prototype of color vision system with three silicon retinas.

One of the interesting issues to be studied in the current chip is its application to color vision. Our silicon retina can be used for each of the three separate channels required for Land's retinex theory of color constancy, and can be used to construct a real-time color vision system. We are now building a prototype of the color vision system consisting of three wide-dynamic-range silicon retinas for red, green, and blue channels (Figure 2). We hope to publish fuller details of this work over the next few months.

2. S. Kameda and T. Yagi, An analog VLSI chip emulating sustained and transient response channels of the vertebrate retina, IEEE Trans. on Neural Networks 14 (5), pp. 1, 2003.
   
   
3. S. Decker, R. D. McGrath, K. Brehmer and C. G. Sodini, A 256×256 CMOS imaging array with wide dynamic range pixels and column-parallel digital output, IEEE J. Solid State Circuits 33 (12), pp. 2, 1998.
   
   
4. E. H. Land and J. J. McCann, Lightness and retinex theory, J. Opt. Soc. Am. 61, pp. 1-11, 1971.
   
   
5. A. Moore, J. Allman and R. Goodman, A real-time neural system for color constancy, IEEE Trans. on Neural Networks 2 (2), pp. 237-246, 1991.