Next: A Foveated Retina Up: Sensor and Processor Previous: An Adaptive Computer

A VLSI Foveated Retina

Authors: [tex2html_wrap4220]R. Wodnicki, G. Roberts, M. D. Levine

Investigator username: levine

Category: perception

Subcategory: sensor and processor design

Machine vision is notorious for its computational expense. In practical applications this often translates into hulking, power hungry systems of distributed processors. Substantial amounts of image information must be transmitted and processed at video rates in order to single out what are typically, small nuggets of useful information. This waste may be tolerated in a controlled laboratory environment, however for useful mobile robotics applications it is unacceptable.

In a world of anthropomorphic machines asked to fit into traditionally human roles, it is perhaps not surprising that the answer to this dilemma is right before our eyes. Nature has solved the problem by splitting the human visual field into two distinct regions. The fovea is a small region of photoreceptors at the center of the human retina used for retrieving detailed image information. Outside of this region, in the periphery, visual acuity decreases in an exponential fashion. Instead of receiving the entire visual field at high resolution and then reducing it to the small subset required for most vision tasks, the brain uses the rudimentary information provided by the periphery to decide which details in the image are of interest. It then instructs the eye to focus these important details on the fovea. This elegant compromise allows for very sophisticated processing without the waste of the traditional machine vision approach.

Motivated by this biological solution, we are developing a foveated imaging device for computer vision in Complimentary Metal Oxide Semiconductor (CMOS) Very Large Scale Integration (VLSI) technology. The device performs a geometric image transformation which has similar spatial sampling properties to those found in the visual pathways of human beings. In this so-called foveated retina, polar coordinates in the image plane are mapped to rectangular coordinates in the cortical plane. Analogous to the human retina, a small region of photosensitive devices at the center of the imaging array, termed the fovea, provides uniform high resolution sensing. Outside of this region, in the periphery, light sensors are arranged radially and increase exponentially in size, resulting in a tremendous saving in data-handling complexity.

Research has concentrated on adapting a typical CMOS fabrication process to this imaging application. As opposed to the traditional approach using Charge Coupled Device (CCD) cameras, this would allow the integration of image capture and analog as well as digital information processing on a single chip. The advantage of such an approach is the facilitation of massively parallel computation, as well as low power consumption, very low system mass and high speeds of operation.



Next: A Foveated Retina Up: Sensor and Processor Previous: An Adaptive Computer


baron@cim.mcgill.ca