| Summary: | 博士 === 國立臺灣大學 === 電機工程學系 === 85 === In this thesis, a neuromorphic vision system which simulates the
neural circuitry (or functions) of the retina and primary visual
cortexof vertebrates was proposed.The proposed silicon retina
has the capability of original image extraction, smoothing,
segmentation, edge detection, and edge enhancement. Pixel
circuit of the silicon retina consists of a parasitic
phototransistor, a Wilson current mirror, two coupling resistors
(implemented by PMOS transistors), two resistive fuses, and
some switches for mode selection.The resistive fuse, exhibiting
independent tunability between its resistance and cutoff
voltage, is responsible for image smoothing andsegmentation. By
suitably biasing some transistors in pixels, totally six modes
of the silicon retina are obtainable,each of these is dedicated
to a specific image processing task. Functionality of the
silicon retina had been experimentally verified.Moreover, a
high-precision median filters, for both voltage and current
mode, was developed to eliminate impulsive-like noise causedby
mismatch between pixel circuits. To achieve real-time median
filtering, a new current sample-and-hold circuit and animage
scanning strategy is proposed. Measurements show that the median
filter and the current sample-and-hold circuit function
properly.Simple feature extraction performed in the primary
visual cortex is realized by a Gaussian/Square similarity
computation circuit. The circuit exhibits independent control
between the width and the peak value of the Gaussian/square
function.Specific edge features can be extracted by performing
template matching throughout the image.Finally, several chips
intended for specific image processing tasks are integrated on
the same substrate for realization of a more complete
neuromorphic vision system.
|
|---|
