| In this experimental research, the signals used to perform Finite Impulse Response image filtering are altered with Error Spectrum Shaping (ESS) so as to preclude corruption by circuit imperfections, in contrast to modifying the circuits to meet some improved level of performance. In using this technique, the representational noise caused by the circuit's resolution inaccuracy is pushed into an unused portion of the signal band, permitting the in-band portion of the signal to be more effectively captured and carried by the noisy analog channels. The suitability of ESS as a method for averting the detrimental effects of the error types inherent in analog array computations is established, with in-band noise reduction through ESS demonstrated for both binary quantization noise and random Gaussian errors. The design, fabrication, and testing of specific CMOS VLSI image convolution circuits are described, and the measured data collected from these circuits is used to model their behavior. The behavioral models are then used in convolution simulations of Nyquist, oversampled, and 1st order and 2nd order noise shaped images of varying resolutions and the SNR improvements quantified. |
