## A One-Dimensional Analog VLSI Implementation for Nonlinear Real-Time Signal Preprocessing (2001)

Citations: | 2 - 1 self |

### BibTeX

@MISC{Wiehler01aone-dimensional,

author = {K. Wiehler and J. Heers and C. Schno Rr and H. S. Stiehl and R. -r. Grigat},

title = {A One-Dimensional Analog VLSI Implementation for Nonlinear Real-Time Signal Preprocessing},

year = {2001}

}

### OpenURL

### Abstract

this paper (see Theory section) can be restated in the CNN framework [25] as well. The model has to be altered to the hardwarefriendly full-range model [26] using nonlinear feedback templates. Nevertheless the CNN framework is defined on a discrete grid and therefore limited in its applications. Various CMOS VLSI implementations of CNNs have been reported [27]. They can be discriminated by fixed [28] or variable [29, 30] templates, current [26] or voltage mode [31], full [32] or limited [33] CNN-model, on-chip [34] or o#-chip sensors. They are referenced here in detail as a starting point for further investigations in the field of analog VLSI systems for signal processing.

### Citations

3895 |
Stochastic relaxation, gibbs distributions, and the bayesian restoration of images
- Geman, Geman
- 1984
(Show Context)
Citation Context ...e survey [7]) being superior to ad-hoc smoothing schemes. A common problem with these approaches, however, is their computational complexity from an optimization pointof-view. Stochastic optimization =-=[4] is not -=-feasible for typical image sizes, and deterministic annealing procedures [5, 6] cannot guarantee to obtain a ‘‘good’’ local minimum. Therefore, the use of non-quadratic but convex functionals ... |

1549 |
The finite element methods for elliptic problems
- Ciarlet
- 1978
(Show Context)
Citation Context ...lying continuous problem formulation, like the rotational invariance of Figure 2. The characteristic function r(t).sthe smoothness term, for example. For a sound introduction to FEM we refer to, e.g. =-=[36]-=-. The basic idea of the Finite Element Method is the restriction of optimization problems to finite-dimensional subspaces. Let { ,..., n} be basis functions of a finite-dimensional subspace Hh H. Then... |

1448 |
Nonlinear total variation based noise removal algorithms
- Rudin, Osher, et al.
- 1992
(Show Context)
Citation Context ...called tiny-tanh network is a further development of Sivilotti’s [20] nonlinear saturating resistor. By introducing a high-gain positive feedback, the network resembles a discrete TV-norm realizatio=-=n [21]-=-. Nevertheless, the implementation lacks both an efficient input-output mechanism and the possibility for controlling the regularization parameters. Most recent implementations of resistive fuse algor... |

957 |
Optimal approximation by piecewise smooth functions and associated variational problems
- Mumford, Shah
- 1989
(Show Context)
Citation Context ...equations, analog VLSI implementations allow for real-time application, i.e. at (video) frame rates. The most powerful, in the sense of general, variational approach was suggested by Mumford and Shah =-=[1]-=-. The theory of this model (e.g. properties of solutions) are subject of an intensive mathematical research and, with focus on efficient algorithms, a lot of work has to be done. A drawback of this co... |

751 | Visual Reconstruction
- Blake, Zisserman
- 1987
(Show Context)
Citation Context ...e approaches, however, is their computational complexity from an optimization pointof-view. Stochastic optimization [4] is not feasible for typical image sizes, and deterministic annealing procedures =-=[5, 6] cannot -=-guarantee to obtain a ‘‘good’’ local minimum. Therefore, the use of non-quadratic but convex functionals has been advocated to simplify image smoothing from a computational viewpoint [2,8,9]. ... |

665 |
Analog VLSI and Neural Systems
- Mead
- 1989
(Show Context)
Citation Context ...hardware concepts allow to map classes of early vision problems onto time-continuous high-speed analog circuitry by the direct use of voltage, current and charge relationships of the physical devices =-=[12]-=-. The behavioral characteristic of electronic circuits as well as the solution of variational equations can be described by nonlinear differential equations [2]. An analog circuit can solve the equati... |

245 |
Cellular neural networks: theory
- Chua, Yang
- 1988
(Show Context)
Citation Context ...can be found in [22]. Also, in the last decade a lot of research in the field of massively-parallel hardware has been triggered by the Cellular Neural Network (CNN) theory introduced by Chua and Yang =-=[23, 24]-=-. A CNN is defined on a timecontinuous, discrete grid and consists of an uniform array of analog nonlinear dynamic computing cells. They are interconnected in a local neighborhood and RTI: 19990218 6.... |

96 |
Computational vision and regularization theory. Nature 317
- Poggio, Torre, et al.
- 1985
(Show Context)
Citation Context ...citors and the finite power dissipation [13]. This is the major motivation for the design of analog VLSI networks suitable for solving such early vision problems. Inspired by biology, Koch and Poggio =-=[14,15] i-=-nvestigated the relationship between mathematically stated regularization schemes and analog networks.sKirchhoff’s current and voltage laws, which represent conservation and continuity restrictions ... |

92 |
Cellular neural networks: Applications
- Chua, Yang
- 1988
(Show Context)
Citation Context ...can be found in [22]. Also, in the last decade a lot of research in the field of massively-parallel hardware has been triggered by the Cellular Neural Network (CNN) theory introduced by Chua and Yang =-=[23, 24]-=-. A CNN is defined on a timecontinuous, discrete grid and consists of an uniform array of analog nonlinear dynamic computing cells. They are interconnected in a local neighborhood and RTI: 19990218 6.... |

81 |
A common framework for image segmentation
- Geiger, AL
- 1991
(Show Context)
Citation Context ...e approaches, however, is their computational complexity from an optimization pointof-view. Stochastic optimization [4] is not feasible for typical image sizes, and deterministic annealing procedures =-=[5, 6] cannot -=-guarantee to obtain a ‘‘good’’ local minimum. Therefore, the use of non-quadratic but convex functionals has been advocated to simplify image smoothing from a computational viewpoint [2,8,9]. ... |

51 |
VLSI analogs of neuronal visual processing: a synthesis of form and function
- Mahowald
- 1992
(Show Context)
Citation Context ...to describe prerequisites on the circuit level which were originally defined on higher mathematical levels. For the realization of those networks, substantial work has been carried out by Mead et al. =-=[12, 16, 17]-=- by utilizing sub-threshold standard MOS technology for massively-parallel analog signal processing. Good examples are optical motion sensors, silicon retinas, etc. By including optical sensors (CMOS-... |

35 | Analog VLSI signal processing: Why, where and how
- Vittoz
- 1994
(Show Context)
Citation Context .... Furthermore, the operation mode, weak or strong inversion, has to be taken into account. Strong inversion circuits offer better accuracy and higher speed compared to circuits in weak inversion mode =-=[39]-=-. Nevertheless, the latter ones have a bipolarlike transfer characteristic which could be used to build efficient circuits for special applications (translinear circuits [40]). On the layout level car... |

33 |
Discontinuity preserving regularization of inverse visual problems
- Stevenson, Schmitz, et al.
- 1994
(Show Context)
Citation Context ...es [5, 6] cannot guarantee to obtain a ‘‘good’’ local minimum. Therefore, the use of non-quadratic but convex functionals has been advocated to simplify image smoothing from a computational vi=-=ewpoint [2,8,9]-=-. Furthermore, although being much simpler, convex functionals nevertheless provide reasonable approximations (cf. [10]) to the prototypical but mathematically and computationally sophisticated variat... |

23 |
Ill-posed problems in early vision: from computational theory to analogue networks
- Poggio, Koch
- 1985
(Show Context)
Citation Context ...citors and the finite power dissipation [13]. This is the major motivation for the design of analog VLSI networks suitable for solving such early vision problems. Inspired by biology, Koch and Poggio =-=[14,15] i-=-nvestigated the relationship between mathematically stated regularization schemes and analog networks.sKirchhoff’s current and voltage laws, which represent conservation and continuity restrictions ... |

22 | Translinear circuits in subthreshold
- Andreou, Boahen
- 1996
(Show Context)
Citation Context ...ts in weak inversion mode [39]. Nevertheless, the latter ones have a bipolarlike transfer characteristic which could be used to build efficient circuits for special applications (translinear circuits =-=[40-=-]). On the layout level careful placement and routing techniques have to be used to reduce mismatch as far as possible [41]. By introducing the physical normalization values Iunit, Vunit and gunit Iun... |

16 |
A two-dimensional analog VLSI circuit for detecting discontinuities
- Harris, Koch, et al.
- 1990
(Show Context)
Citation Context ...ed in a mechanistic way, boundary conditions are incorporated automatically, and the discrete approach obtained is consistent in the sense that, under certain conditions [2], discrete solutions uh to =-=(3)-=-, for example, converge to the continuous solution u for vanishing mesh width. In this way discrete versions of our approach allow to maintain favorable properties of the underlying continuous problem... |

16 |
Resistive fuses: analog hardware for detecting discontinuities in early vision
- Harris, Koch, et al.
- 1989
(Show Context)
Citation Context ...tible photo diodes and transistors) vision systems can be implemented on one single chip. For current successful implementations see, for example, [18]. Build upon Mead’s fundamental research, Harri=-=s [3, 19]-=- implemented nonlinear resistive networks for early vision tasks. Utilizing resistive fuse elements he implemented the non-convex approach of Blake and Zissermann [5] (weak membrane model). He achieve... |

14 |
Analog VLSI Integration of Massive Parallel Signal Processing Systems
- Kinget, Steyaert
- 1997
(Show Context)
Citation Context ...amework is defined on a discrete grid and therefore limited in its applications. Various CMOS VLSI implementations of CNNs have been reported [27]. They can be discriminated by fixed [28] or variable =-=[29, 30]-=- templates, current [26] or voltage mode [31], full [32] or limited [33] CNN-model, on-chip [34] or off-chip sensors. They are referenced here in detail as a starting point for further investigations ... |

13 |
Optimizing MOS transistor mismatch
- Lovett, welten, et al.
- 1998
(Show Context)
Citation Context ...el of the nonlinear transfer characteristic (see Figure 5). The error should be as small as possible and independent of the data input and output. Random errors are caused by technology imperfections =-=[37]-=-. Due to this random influence even equal transistors in the schematic behave slightly different in their corresponding realization (mismatch). Moreover one can discriminate between intra- and inter-d... |

12 |
Smart-pixel cellular neural networks in analog current-mode CMOS technology
- Espejo, Rodríguez-Vázquez, et al.
- 1994
(Show Context)
Citation Context ...mplementations of CNNs have been reported [27]. They can be discriminated by fixed [28] or variable [29, 30] templates, current [26] or voltage mode [31], full [32] or limited [33] CNN-model, on-chip =-=[34]-=- or off-chip sensors. They are referenced here in detail as a starting point for further investigations in the field of analog VLSI systems for signal processing. Contribution and organization In our ... |

11 |
Real-time visual computations using analog CMOS processing arrays
- Sivilotti, Mahowald, et al.
- 1987
(Show Context)
Citation Context ...mbrane model). He achieved a robust solution by a graduated convexity mechanism which is inherent to his resistive fuse circuit. The socalled tiny-tanh network is a further development of Sivilotti’=-=s [20]-=- nonlinear saturating resistor. By introducing a high-gain positive feedback, the network resembles a discrete TV-norm realization [21]. Nevertheless, the implementation lacks both an efficient input-... |

10 |
A programmable analog cellular neural network CMOS chip for high speed image processing
- Kinget, Steyaert
- 1995
(Show Context)
Citation Context ...amework is defined on a discrete grid and therefore limited in its applications. Various CMOS VLSI implementations of CNNs have been reported [27]. They can be discriminated by fixed [28] or variable =-=[29, 30]-=- templates, current [26] or voltage mode [31], full [32] or limited [33] CNN-model, on-chip [34] or off-chip sensors. They are referenced here in detail as a starting point for further investigations ... |

10 |
Cellular Neural Networks: Theory and Circuit Design
- Nossek, Seiler, et al.
- 1992
(Show Context)
Citation Context ...re limited in its applications. Various CMOS VLSI implementations of CNNs have been reported [27]. They can be discriminated by fixed [28] or variable [29, 30] templates, current [26] or voltage mode =-=[31]-=-, full [32] or limited [33] CNN-model, on-chip [34] or off-chip sensors. They are referenced here in detail as a starting point for further investigations in the field of analog VLSI systems for signa... |

9 |
Current-Mode Techniques for the Implementation of Continuous and Discrete-Time Cellular Neural Networks
- Rodríguez-Vázquez, Espejo, et al.
- 1993
(Show Context)
Citation Context ...ions of the regularization equation proposed in this paper (see Theory section) can be restated in the CNN framework [25] as well. The model has to be altered to the hardwarefriendly full-range model =-=[26]-=- using nonlinear feedback templates. Nevertheless the CNN framework is defined on a discrete grid and therefore limited in its applications. Various CMOS VLSI implementations of CNNs have been reporte... |

8 |
A 16x16 Cellular Neural Network Universal Chip: the First Complete Single-chip Dynamic Computer Array with Distributed Memory and with
- Cruz, Chua
- 1998
(Show Context)
Citation Context ...in its applications. Various CMOS VLSI implementations of CNNs have been reported [27]. They can be discriminated by fixed [28] or variable [29, 30] templates, current [26] or voltage mode [31], full =-=[32]-=- or limited [33] CNN-model, on-chip [34] or off-chip sensors. They are referenced here in detail as a starting point for further investigations in the field of analog VLSI systems for signal processin... |

5 |
A current-mode cellular neural networks implementation
- Varrientos, Sánchez-Sinencio, et al.
- 1993
(Show Context)
Citation Context ...heless the CNN framework is defined on a discrete grid and therefore limited in its applications. Various CMOS VLSI implementations of CNNs have been reported [27]. They can be discriminated by fixed =-=[28]-=- or variable [29, 30] templates, current [26] or voltage mode [31], full [32] or limited [33] CNN-model, on-chip [34] or off-chip sensors. They are referenced here in detail as a starting point for fu... |

4 |
Performance comparison of analog and digital circuits
- Hosticka
- 1985
(Show Context)
Citation Context ...near differential equations [2]. An analog circuit can solve the equations in fractions of a second. The speed is limited only by the unavoidable parasitic capacitors and the finite power dissipation =-=[13]-=-. This is the major motivation for the design of analog VLSI networks suitable for solving such early vision problems. Inspired by biology, Koch and Poggio [14,15] investigated the relationship betwee... |

3 |
Entwurf und Analyse nichtlinearer Diffusionsverfahren fu¨r die Bild-verarbeitung
- Sprengel
- 1996
(Show Context)
Citation Context ...see Figure 2) characterizes the adaptive smoothing process (1). Partial integration shows that, at least formally, u may be considered as the steady-state solution of the nonlinear diffusion equation =-=[2, 35]: du r -=-ru dt jj ðÞru ð Þ ðugÞ, ð4Þ with Neumann boundary conditions. Discretization For the discretization of variational problems, FEM is the first choice. FEM can be applied in a mechanistic way, b... |

2 |
Convex energy functionals in the DA model
- Li, Huang
- 1995
(Show Context)
Citation Context ...es [5, 6] cannot guarantee to obtain a ‘‘good’’ local minimum. Therefore, the use of non-quadratic but convex functionals has been advocated to simplify image smoothing from a computational vi=-=ewpoint [2,8,9]-=-. Furthermore, although being much simpler, convex functionals nevertheless provide reasonable approximations (cf. [10]) to the prototypical but mathematically and computationally sophisticated variat... |

2 |
On globally asymptotically stable continuous-time CNNs for adaptive smoothing of multidimensional signals
- rr, Stiehl, et al.
- 1996
(Show Context)
Citation Context ...nd therefore are best suited for direct analog implementation. Discretized formulations of the regularization equation proposed in this paper (see Theory section) can be restated in the CNN framework =-=[25]-=- as well. The model has to be altered to the hardwarefriendly full-range model [26] using nonlinear feedback templates. Nevertheless the CNN framework is defined on a discrete grid and therefore limit... |

1 |
Unique reconstruction of piecewisesmooth images by minimizing strictly convex non-quadratic functionals
- rr, C
- 1994
(Show Context)
Citation Context ...re, as yet, an analog VLSI implementation seems to be impossible to achieve. A class of variational approaches which are less powerful but have globally favorable properties was proposed by Schno¨ rr=-= [2]-=- (see the Problems statement section). As a special case this class subsumes a continuously formulated approach which resembles an approach by Harris et al. [3], suggested for an analog implementation... |

1 |
A study of a convex variational approach for image segmentation and feature extraction
- rr, C
- 1998
(Show Context)
Citation Context ...has been advocated to simplify image smoothing from a computational viewpoint [2,8,9]. Furthermore, although being much simpler, convex functionals nevertheless provide reasonable approximations (cf. =-=[10]-=-) to the prototypical but mathematically and computationally sophisticated variational smoothing approach of Mumford and Shah [1]. Despite of efficient digital implementations of numerical schemes for... |

1 |
Parallel and globally convergent iterative numerical schemes for computationally efficient nonlinear variational image segmentation
- Heers, rr, et al.
- 1999
(Show Context)
Citation Context ...matically and computationally sophisticated variational smoothing approach of Mumford and Shah [1]. Despite of efficient digital implementations of numerical schemes for solving variational equations =-=[11]-=-, the computational effort prohibits applications for which frame-rate performance is required. As an alternative, analog hardware concepts allow to map classes of early vision problems onto time-cont... |

1 |
Investigations of Visual Transduction and Motion Processing
- Delbrueck
- 1993
(Show Context)
Citation Context ...to describe prerequisites on the circuit level which were originally defined on higher mathematical levels. For the realization of those networks, substantial work has been carried out by Mead et al. =-=[12, 16, 17]-=- by utilizing sub-threshold standard MOS technology for massively-parallel analog signal processing. Good examples are optical motion sensors, silicon retinas, etc. By including optical sensors (CMOS-... |

1 |
Current mode implementation of a neural algorithm for image preprocessing
- bler, Werner, et al.
- 1996
(Show Context)
Citation Context ...etinas, etc. By including optical sensors (CMOS-compatible photo diodes and transistors) vision systems can be implemented on one single chip. For current successful implementations see, for example, =-=[18]. -=-Build upon Mead’s fundamental research, Harris [3, 19] implemented nonlinear resistive networks for early vision tasks. Utilizing resistive fuse elements he implemented the non-convex approach of Bl... |

1 |
A floating gate MOS implementation of resitive fuse
- Sawaji, Nagai, et al.
- 1998
(Show Context)
Citation Context ...s. Most recent implementations of resistive fuse algorithms utilizing floating-gate MOS-transistors for both, controlling the transfer characteristic and long-term error compensation, can be found in =-=[22]-=-. Also, in the last decade a lot of research in the field of massively-parallel hardware has been triggered by the Cellular Neural Network (CNN) theory introduced by Chua and Yang [23, 24]. A CNN is d... |

1 |
Fast and compact 16 by 16 cellular neural network
- Piasio, Dawidwiziuk, et al.
- 1997
(Show Context)
Citation Context ...ons. Various CMOS VLSI implementations of CNNs have been reported [27]. They can be discriminated by fixed [28] or variable [29, 30] templates, current [26] or voltage mode [31], full [32] or limited =-=[33]-=- CNN-model, on-chip [34] or off-chip sensors. They are referenced here in detail as a starting point for further investigations in the field of analog VLSI systems for signal processing. Contribution ... |

1 |
A 0.8 mm CMOS 2-D programmable mixed-signal focal-plane array processor with onchip binary imaging and instructions storage
- nguez-Castro, Espejo, et al.
- 1997
(Show Context)
Citation Context ...l enough to guarantee sufficient bias condition on all chips (yield) for a proper performance. For the case of cascading multiple analog chips an appropriate compensation method has to be implemented =-=[38]-=-.s8 K.WIEHLER ETAL. Table 2. Necessary conditions for reasonable segmentation results. d denotes the global error and s denotes the local error, i.e. the standard deviation of normally distributed off... |

1 |
Auswirkungen von Bauteil-sto¨ rungen in der analogen Signalverarbeitung am Beispiel eines CMOS-Transkonduktanzversta¨ rkers fu¨ r zellula¨ re neuronale Netze
- Wiehler, Klingspohr
(Show Context)
Citation Context ...n relaxation time constants of microseconds for the used capacitors (&1 pF). Based on an analytical error analysis, the simple core circuits were optimized regarding their minimum possible variations =-=[42]-=-. With these results, area and power consumption can be calculated for the specification of the application. The result of a Monte-Carlo simulation for the used nonlinear circuit is depicted in Figure... |

1 |
Large-image CNN hard-ware processing using a time multiplexing scheme
- Gyvez, Wang, et al.
- 1996
(Show Context)
Citation Context .... In this case the signal vector has to be broken apart into overlapping blocks which have to be processed sequentially. As a consequence, the parts have to be merged in a further postprocessing step =-=[43]-=-. Our solution to these problems is based on a dynamic reconfigurable circular architecture which is shifted along the data vector at a very high rate [44]. For the 1D case the neighborhood Nr(i) of t... |

1 | Dynamic circular cellular networks for adaptive smoothing of multi-dimensional signals
- Wiehler, Lembcke, et al.
- 1998
(Show Context)
Citation Context ...o be merged in a further postprocessing step [43]. Our solution to these problems is based on a dynamic reconfigurable circular architecture which is shifted along the data vector at a very high rate =-=[44]. -=-For the 1D case the neighborhood Nr(i) of the cell Ci within a radius r on a grid D is defined by (Figure 10(a)). NrðiÞfCjjminðhi jin,h j iiNÞ r, j 0, ..., N 1, j6 ig i0, ..., N 1 ð14Þ hi N is ... |

1 |
Real-time Adaptive Smoothing with a 1-D Non-linear Relaxation Network in Analogue VLSI Technology
- Wiehler, Grigat, et al.
- 1998
(Show Context)
Citation Context ...ly and efficiently perform computationally intensive nonlinear smoothing of 1D signals. In addition to the 1D experiments, we carried out several experiments with 2D gray-value images (see Figure 16, =-=[45]-=-). The goal of these experiments was to investigate from a practical perspective to which extent even only the 1D chip with its reduced design complexity is capable of satisfactorily smoothing 2D imag... |