Product News: Vision Systems
Vision System has built-in image recognition capability.
Press Release Summary:
December 11, 2001 - NeuroSight board has neural network for pattern learning and recognition so it can be configured as smart camera for person identification, fingerprint recognition, target tracking, wood grading, PCB board-inspection, and similar applications. Image recognition is performed by scanning images and comparing features with examples stored in neural network. Teaching and configuring can be performed through PC, keypads or touch screen panels.
Original Press Release
First Embedded Vision Board With Built-In Image Recognition Capabilities
Press release date: November 20, 2001
Tokyo, November 20th, 2001 - General Vision, manufacturer of smart vision systems, announces at the Embedded Technology Conference & Exhibition 2001 the release of its NeuroSight embedded board for high-speed image recognition. The board is a self-contained system that can acquire video images, recognize known objects in these images and transmit results to the outside world. The key components of the NeuroSight are a CMOS sensor for video acquisition, a ZISC (Zero Instruction Set Computing) silicon neural network for pattern learning and recognition and an FPGA (Field Programmable Gate Array) that makes NeuroSight programmable for a variety of vision applications. Through this combination of CMOS, ZISC and FPGA, NeuroSight can be configured as a smart camera for person identification, fingerprint recognition, target tracking, wood grading, PCB board-inspection, fruit sorting and more. This is all done without the need of model programming or PC computing power.
The new card integrates on a compact size of 3"x 4" a color CMOS sensor module with a resolution of 352x288 pixels, a network of 156 neurons, a memory bank, an FPGA of 150K gates, and several I/O user lines. NeuroSight is intended for integration in a stand-alone smart camera or other appliance, but it can be interfaced to a PC serial or parallel port for the purpose of configuring the FPGA and loading a recognition engine in the neural network. The FPGA can be programmed to perform operations such as image acquisition control, on-the-fly pixel processing, feature extractions, and communication with the neural network and I/O lines.
The ZISC neural network works like the human brain. Its knowledge is built by learning examples of objects, or rather their feature vectors, and their category (i.e. good or bad; John or Robert or Sam; defect grades from 1 to 10). If taught by a moderate teacher, the engine will have the tendency to over-generalize what it knows and adapt to new objects or situations which were not part of the learning examples. On the other hand, if the teacher is conservative, the engine will be very precise but also narrow-minded. A PC-based graphical interface can be used for the annotation of the objects, the construction of the engine and the validation of its throughput and accuracy. Image recognition is performed by scanning the images per blocks of pixels and comparing their feature vectors with the examples stored in the neural network. If the feature vector of a block is recognized by one or more neurons, the network returns the categories of all firing neurons. The block position and categories can then be conditioned into a format significant to the application and transmitted through the I/O ports of NeuroSight. For example, results can be formatted to report the coordinates of all identified objects, or the individual category of each object, or the most significant category in the image or other application-specific data.
Once an engine delivers satisfactory results, it can be transferred into the neurons of the NeuroSight and its settings loaded on the FPGA. NeuroSight can then be disconnected from the PC and make its own decision. Teaching and configuring the NeuroSight can be performed through a PC system, but also through other appliances such as handheld keypads or industrial touch screen panels with a limited number of controls. At any time a new program can be loaded into the FPGA of the NeuroSight- embedded board and a new recognition engine loaded into its neural network. The FPGA can be programmed directly or through a CPLD and Flash memory. Users who wish to interface with a digital video sensor other than the one supplied with NeuroSight can do so by unplugging the default CMOS sensor module and connecting one of their choice through the 40-pin I/O connector of the NeuroSight.
Pricing information NeuroSight Embedded is available now for a list price of $795 U.S. with a color CIF video sensor (model # NS_EM2_C). The board is delivered with power supply, cables for the RS232 and parallel ports, CDROM with a NeuroSight Loader utility, an HDL library and an example bitstream file for basic object recognition.