|
|
|
“Digital Motion Estimation and Image Restoration on VLSI”
by Ji-Chien Lee
July 1991
The processing of video signals often requires a tremendous computational capability which can only be achieved by using highly parallel processing architectures. The inherent massive parallelism of artificial neural network architecture for flexible information processing provides a new paradigm of video signal processing. This dissertation describes the computational needs of supercomputing neurocomputers for flexible information processing. Two neural network architectures and the efficient VLSI implementation of video signal processors are presented. In the first VLSI architecture, the motion information from a sequence of image data can be estimated through a two-dimensional multiprocessor array in which each processing element consists of an analog neuroprocessor. Massively parallel neurocomputing is done by compact and efficient neuroprocessors. Local data transfer between the neuroprocessors are performed by using analog point-to-point interconnection scheme. Global data communication between the host computer and neuroprocessors is carried out in the digital common bus. A mixed-signal VLSI neural chip that includes multiple neuroprocessors for fast video motion estimation has been designed. Measured results of the programmable synapse, summing neuron, and associated winner-take-all circuitry are presented. Based on the measurement data, system-level analysis on a sequence of real images were conducted. The device mismatch effect of analog synapse cells has been included during system-level analysis. A 1.5 x 2.8-cm2 chip in a 1.2-mm CMOS technology can accommodate 64 velocity-selective neuroprocessors and achieve 83.2 Giga connections per second. The speed-up factor over a Sun-4/75 SPARC-2 workstation is 24,242 for a system with 128 motion estimation chips. In the second architecture, an analog systolic multiprocessor for high-speed image restoration has been developed. For a two-dimensional image, parallel processing is performed in the row direction and pipelined processing is performed in the column direction. The mixed analog/digital design approach is also used for the implementation of the neural-based image restoration system. Local data computation is executed by analog circuitry to achieve full parallelism and to conserve power dissipation. Inter-processor communication is carried out in the digital format to maintain adequate signal strength across the chips boundary and achieve direct scalability in neural network size. A compact and efficient VLSI neural chip which includes multiple neuroprocessors for real-time digital image restoration has been designed. To use output of neuron as an increment/decrement information to control the pixel register, each neuron can process multiple-bit image information. A 8.0 x 6.0-mm2 chip from a 1.2-mm CMOS technology can accommodate 5 neuroprocessors and the speed-up factor over the Sun-4/75 SPARC-2 workstation is 475. This chip achieves 21.6 Giga connections per second. The future powerful and cheap supercomputing workstations will include flexible information processing capability for our daily lives and scientific applications.