| 
Imager and Intelligence on a Chip ColorCam Chip Prototype, Iguana Robotics, Inc. The objective of this research is to combine generally useful image processing algorithms and an imager device on a single piece of silicon. The result will be a device that will be very inexpensive to produce.
To understand the motivation behind the chip, it is necessary to mention a few facts about information processing in the brain. In the early nineties, neuroscientists and cognitive psychologists were perplexed by the fact that the brain seemed to process visual information along parallel channels. One channel, which funnels information from the back of one's head to a brain region behind the ear, is called the ventral stream. This information flows to areas that are expert at recognizing what a thing is, but strangely, this region did not realize where the object was in space. That function is reserved for a brain region which channels information to a place directed toward the top of one's head, the dorsal processing stream. This flew in the face of then current machine vision object techniques, which first found an object then decomposed it into visual segments.
In the early nineties, a new class of vision recognition algorithms was developed based on gathering statistics from each point in the image. Geometric information, i.e. the relationship between points in the image, was thrown away. How does this affect chip design? It turns out that this class of image processing is well suited to EBT techniques On-chip processing is laid out as a mixed signal pipeline. The processing rate using this technique easily exceeds that needed for object recognition or color segmentation tasks.
Image Acquired with Chip
Information is acquired with a color pixel array. Currents, representing the R,G,B (red blue and green channels) luminosity values are scanned out sequentially and normalized by the sum (R+G+B) of the 3 channels. This information is then passed to a color converter module. This converter performs a nonlinear transformation from RGB values to Hue-Saturation-Intensity (HSI) values. HSI is a color system that mimics how artists express color. This nonlinear transformation is performed by a mixed signal circuit making use of a table lookup to compute a transcendental function needed for the RGB->HSI conversion. HSI is a good space for operations such as image segmentation and histogram matching. For example, one property of the HSI representation is that it discounts shadows. That is, theoretically, Hue and Saturation components will ignore illumination variation. While this is clearly a desirable property, it is difficult to compute Hue and Saturation accurately in practice using a conventional digitization and image processing system. For various technical reasons, we can avoid problems inherent in digital systems which lead to instabilities in the Hue computation component. Once the HSI transformation is computed, we then can match the H and S components to predefined 'targets.' If the incoming signal falls with in an epsilon region of the center of the targets color, a pixel level match occurs and an 'match' bit is set. This bit is then transferred to a segmentation circuit to mask RGB readout, as well as used to toggle a counter array, and thus compute the histogram of an entire image or a section of an image. Once the histogram is computed, it can be matched versus a stored template.
Example of color Histogramming. In this case, a stock digital image is downloaded to chip for historgram analysis.
Example of Segmentation. Upper left: Original Image. Upper right: segmented into 32 levels. Lower Left: Greenish color issolated. Lower Right bluish color issolated. Image is acquried with standard sensor and downloaded to chip .
In general, 100% recognition rate is not possible using a color histogramming alone. However, extremely good recognition rates can be achieve by using a cooperative target. For example, by using objects coded with a color 'key.' Using the concept of a cooperative target, we can tune the chip to recognize a class of targets.
What are the uses for this chip? We see the chip being embedded as part of an intelligent toy. The chip would allow the recognition of a set of objects associated with a toy. For example, a toy soldier might have assorted jet airplanes, a tanks, other soldiers. The soldier might also have a specially constructed landscape. The toy soldier would be able to scan the landscape and recognize targets, triggering sounds or actions.
The color segmentation ability could be used to segment skin tones as the front end of an intelligent camera. Finally, we are currently developing this chip with an eye on retinal implant technology. For example, the same technology which would allow the chip to recognize objects associated with a toy would allow a blind person to 'see' specially coded objects in the environment (key, doors knobs, traffic light) by way of retinal implants, an emerging technology.
See Demo video from 2002 on Download Page
Take me to the DownLoad Page
|
