CCD sensors mimic the eye
Most commercially available CCD image sensors use arrays of square, regularly spaced sensors to capture images. Although this results in visually acceptable images with linear resolution, the amount of data generated can overwhelm all but the most sophisticated parallel processors.
In pattern-recognition applications especially, visual sensors that mimic the human retina can reduce the amount of data while retaining a high resolution and wide field of view. Such space-variant devices known as foveated sensors have been developed at the University of Genoa (Genoa, Italy) in collaboration with IMEC (Belgium) using CCD and CMOS technologies.
"Foveated vision reduces the amount of processing required and lends itself to image-processing and pattern-recognition tasks that are currently performed with uniformly spaced imagers," says professor Giulio Sandini of the University of Genoa. Such devices closely match the way human beings focus on images and may find use in pattern-recognition systems.
"In robotic systems," says Sandini, "the low-resolution periphery of the fovea locates areas of interest and can direct the sensor to them." In the CCD design built by Sandini, the IC has a central high-resolution rectangular region and successive circular outer layers with decreasing resolution.
In the circular region, the IC implements a log-polar mapping of Cartesian coordinates to provide scale- and rotation-invariant transformations. The prototype sensor comprises pixels arranged on 30 concentric circles, each with 64 photosensitive sites. Pixel sizes increase from 30 ¥ 30 µm at the inner circle to 412 ¥ 412 µm at the periphery. With a video rate of 50 frames per second, the CCD sensor generates images with 2 kbytes per frame. According to Sandini, this allows the device to perform tasks such as computation of the time to impact of a target approaching the device. Sandini has also developed a CMOS device with 128 pixels per ring (instead of 64) and 56 rings; plans are underway for a color version.
For more information, contact Sandini via e-mail at giulio@dist. unige.it.