We fabricated a novel vision system consisting of an analog very large scale integrated(VLSI)vision sensor,namely a silicon retina,and field programmable gate array(FPGA)circuits.The silicon retina executes Laplacian-Gaussian-like spatial filtering and subtraction of consecutive frames with its built-in analog integrated circuit.The output images of the silicon retina are fed into the FPGA circuits that perform further image processing using parallel circuit architecture.Using this system,contours of objects and their direction of motion were extracted within a single-frame sampling-time of the silicon retina(<20ms in indoor illumination)with low computation cost(<10k gates of FPGA circuits).The system is compact and has low power consumption(<2.75W),and is therefore suitable for robotic vision.
We performed high-dynamic-range image reduction method with continuous luminance properties by combining images with long and short exposure times.We then used a spatial lter to make a palameter that retains the outline contrast and applied it to a compressed image with a high dynamic range.The retention mask,which consists of a combination of the spatial lter and the original image,retains the low- and high-illuminance sections of the outline.Therefore,unlike with the conventional arti cial retinal processing algorithm,we can retain contrast without outline reinforcement.We achieved image reduction method of high dynamic range with continuous luminance properties without loss of contrast.
Conventional demosaicing methods often generate an artifact noise near a color edge because they cannot adapt to the changes of colors there. We have therefore proposed a new demosaicing method that interpolates a missing color on the basis of results obtained from an analysis of color changes in a local region. We propose analyzing the correlation of color signals by comparing the differences in color signals and by using regression analysis. The missing color is thus interpolated adaptively. We have compared conventional demosaicing methods with our method and have shown that our method can be used to reduce the artifact noise and give a higher quality image.
By transferring electric charges in a vertical register of a CCD imager,we have made a new imagestabilizing method.Furthermore,we have measured its effect.A small digital still camera(DSC),such as the camera module that is employed in mobile phones,needs image-stabilizing means that are both low-cost and durable.With this method,multiple short exposures of an object are taken and electric charges in the vertical register are added to make non-blurring still pictures.The first exposure in the vertical register is transferred upward or downward until the next one is read out in the same direction of a moving optical image in order to automatically add the same object.Miniaturized cameras can thus be made much more cheaply than previously because flame memory and calculations are not needed to detect motion vectors.
A wide dynamic range 64x64 CMOS image sensor with 20x20um 2 pixel that combines a lateral-overflow integration voltage-readout operation with the current readout operation from the buried photo-diode hasbeen developed.In the voltage readout operation,an over 160-dB dynamic range image with a linear response isobtained from sequential electronic shutter operations from 1/30 s to 1/130 ks with a dynamic range of about 100-dB within a light intensity range from about 10
-2lx to 106lx.In addition,an over 200-dB dynamic range perform-ance of up to 108lx or more with few time exposures in the high illuminance region is made possible by combiningit with the current reading operation.More than 40 dB for all noises,including photon shot noise,occurs around all the switching points.
We present very wide dynamic range (WDR) CMOS image sensor that integrates improved high-speed column-parallel cyclic ADCs.The proposed signal readout technique of extremely short accumulation (ESA) enables the dynamic range to be expanded to a very high illumination region.Including the ESA signals,a total of four different accumulation time signals are read out in one frame period using a burst readout technique.To achieve the high-speed and high-quality signal readout required for multiple exposure signals,column parallel high-speed A/D converters are integrated at the upper and lower sides of the pixel arrays.This improved cyclic ADC,with reduced random noise,better linearity and offset deviation,expands the dynamic range in the low illumination region.The resulting dynamic range is maximally dB.The improved 12-bit cyclic ADC has a differential non-linearity (DNL) of 0.3LSB.
We investigated parallel image compression circuits suitable for integration in high-speed CMOS image sensors.We compared the coding efficiency and hardware complexity of several image compression algorithms that use 2-D DCTs using simulation and logic synthesis,and found that using 4$\ imes$4 point 2-D DCT-based coding methods reduced hardware complexity and improved coding efficiency.We developed a parallel processing architecture for on-sensor image compression that use a processing element array and a data-buffering scheme for parallel data-output.We constructed a prototype 256 $\ imes$ 256 pixel high-speed CMOS image sensor chip that integrates 16 image compression-processing elements and uses 0.25-$\mu$m CMOS technology.The area of the image compression circuits is 80\% of the image array with 15 $\mu$m square pixels.The entire chip could be operated at a clock frequency of 53.6 MHz,and high-speed images compressed by a factor of four could be read out at 10,000 fps using a 32-bit parallel bus.
Wide dynamic-range image sensors are now used for various applications such as intelligent transport systems and security systems. We propose a new image-sensor that controls the reset timing of the photodiode in each pixel. When a pixel value exceeds a threshold during the integration period, this sensor immediately resets the photodiode and records the number of resets in a memory. By using this sensor, linear-wide dynamic-range images can be obtained.
We developed a smart image-sensor especially for extracting statistical values that contain areas,positions,orientations and shapes of the objects in an image.These values are called moments.The sensor has a small processing circuit in each pixel in order to achieve high resolution,but it outputs various information at real-time.We describe the implementation and performance evaluation of a 128 128 pixel image-moment sensor.
We made an active-matrix high-efficiency electron emission device (HEED) that had scanning driver circuits fabricated with conventional MOS planer technology.A 1-in 256x192 pixel image sensor,consisting of the active-matrix HEED and a high-gain avalanche rushing amorphous photoconductor (HARP) target,was fabricated and tested as a step toward the development of ultrahigh-sensitivity compact image sensors with a large number of pixels.The experimental results show that the prototype sensor can be used to obtain clear images even under dark lighting conditions,such as in moonlight,and that the active-matrix HEED is a promising highspeed image pick-up device.
We carried out a categorical color naming experiment using young adults and elderly people as observers.The experiment was designed to investigate the effects of mode of color appearance (surface and aperture modes) ,and illuminants (four types of fluorescent:D-EDL-D65,N-EDL,EX-L and W) .The experiment used a series of Munsell chips which were presented separately,and the observers were asked to categorize them into one of 11 categories (red,pink,orange,yellow,green,blue,purple,brown,white,gray,and black) .The constancy of the categorization results was high for the young adults but not for the elderly observers.The elderly observers categorization constancy was poor for chips with hues RP,R,YR,and Y.Illuminants did not affect the categorization for the young adults very much but had a relatively large effect on the elderly observers.The color appearance mode had a slight affect on the categorization constancy of both groups of observers.