One-dimensional discrete-time CNN with multiplexed template-hardware

Abstract

This paper presents a novel discrete-time and fully programmable cellular neural network (CNN) suitable for processing one-dimensional (1-D) signals. As 1-D signals are typically very long sequences, the system consists of a linear analog shift register for data I/O coupled to a 1/spl times/n CNN array. In addition to the 1-D CNN architecture, a unique feature of our implementation is that the number of multipliers needed to implement both CNN templates has been minimized. This is conceivable because the multipliers are multiplexed between the A*y and B*u products during alternating phases of the controlling clock. The CNN system has been implemented in current mode based on the S/sup 2/I technique using MOSIS Orbit 2 /spl mu/m CMOS technology. The paper presents a thorough behavioral analysis of the new architecture, circuit-level implementations, and corresponding measured experimental results.