An image is regarded as a 2-D array of pixels and is processed by a 2-D array architecture. The image can be acquired in the usual manner by a raster scan method which produces a 1-D array of pixels at real-time video rates. Two 2-D systolic arrays for a 2-D convolver are presented. They have an architecture which accepts this 1-D array of pixels and processes them in a 2-D array of simple processors. This high degree of parallelism is achieved through matrix-vector formulations of 2-D convolution. One array has a serial input, a serial output, and uses a minimum number of multipliers: the other array has parallel input, parallel output, and is suitable for high-speed processing using slow processing elements. Both arrays are modular with nearest-neighbor communications are are suitable for VLSI implementation. In addition, an algorithm for 2-D convolution that explicitly takes into account the boundary conditions is presented. This feature allows a large image to be partitioned so that each partition may be processed by independent 2-D convolvers. It is then possible to process only a specified section of the image or carry out high-speed parallel processing using as many 2-D convolvers as are available. >
[1]
H. T. Kung.
Special-Purpose Devices For Signal And Image Processing: An Opportunity In Very Large Scale Integration (VLSI)
,
1980,
Optics & Photonics.
[2]
H. T. Kung,et al.
Warp as a machine for low-level vision
,
1985,
Proceedings. 1985 IEEE International Conference on Robotics and Automation.
[3]
Okan K. Ersoy.
Semisystolic Array Implementation of Circular, Skew Circular, and Linear Convolutions
,
1985,
IEEE Transactions on Computers.
[4]
H. T. Kung,et al.
One-Dimensional Systolic Arrays for Multidimensional Convolution and Resampling
,
1984
.