architectures implementing the same CNN can be quite different, but can be predicted theoretically. Gabor filters are preprocessing stages for many image processing and computer vision applications. Unfortunately, they are computationally intensive on a digital computer. Although an analog VLSI chip for Gabor filtering could relieve this bottleneck by computing the filter outputs with less power and in less time than required by serial digital computers, one drawback is a loss in accuracy due to the limited precision with which circuit components can be implemented. This paper describes an analysis of several different possible circuit implementations of an analog VLSI cellular neural network for Gabor filtering which shows that the effect of variations in circuit components can be minimized by proper circuit design.

Abstract-We analyze distortion in analog VLSI arrays for linear image filtering due to component mismatch and nonlinearity. The analysis can be used to evaluate different circuit architectures implementing the same computation for both one-dimensional (1-D) and two-dimensional (2-D) rectangularly and hexagonally sampled arrays. In addition, it can determine which elements within a given architecture are most critical for accurate computation of the output. Based upon the concept of equivalent input sources, it offers both mathematical and intuitive explanations of the responses of the networks to mismatch and nonlinearity.

No abstract

No abstract