We describe a salient point detection model that incorporates recent findings about the response properties of neurons in cortical area V2. It has been reported that certain V2 neurons encode combinations of orientations. The model reported here integrates this finding as the last stage of a hierarchical architecture modeled after the visual processing in mammalian brains. Our tests of this model on real images demonstrate its effectiveness in capturing the intuitive notion of saliency as well as its robustness against viewpoint changes.
An algorithm for fast and robust face tracking with the CNN universal machine is proposed in this paper. It is applied to a driving mechanism for a wheelchair with an on-chip implementation. A novel object tracking CNN visual algorithm is introduced and employed in the tracking of multiple face features. The speed and robustness of this method are achieved due to the parallelism in the visual algorithm, and the tracking of multiple face features. The tracking algorithm is designed to achieve a high frame rate and exploit the specific properties of face features. The face tracking method proposed here was implemented on a Bi-I stand-alone cellular vision system and applied to a wheelchair driving mechanism. The template operations were trained and/or fine-tuned in order to generate chip-specific robust templates. In order to improve performance in environments with varying illumination, an adaptive image capture procedure was also introduced. Our simulations with a 3D model wheelchair showed that the final algorithm is capable of performing tracking with a frame rate of 92 frames/sec, which is supposedly enough for real-time driving in most of the real life situations
We demonstrate a modular hardware architecture for real-time simulation of the outputs of retinotopic arrays of cortical columns, which are simultaneously selective along multiple feature dimensions including retinal position, spatial and temporal frequency, orientation, disparity, and direction of motion
We give the necessary and sufficient conditions for a one-dimensional discrete-time autonomous binary cellular neural networks to be stable in the case of fixed boundary. The results are complete generalization of our previous one [16] in which the symmetrical connections were assumed. The conditions are compared with some stability conditions so far known.
The simulation software COGNI simulates the pattern recognition neural network neocognitron of K. Fukushima (1982). Due to its complexity, simulations can be carried out only on relatively powerful computer systems which are capable of high speed numeric processing and graphic display. There are two versions available, using the IBM PC-AT and the mu VAX II. Neocognitron is able to learn without a teacher. The response of the last layer in forward (afferent) paths is not affected by the pattern's position or by a small change in the shape or size of the stimulus pattern. Even stimuli corrupted with noise are successfully recognized. The autoassociation is also achieved in the last layer of backward (efferent) paths i.e. in the autoassociation plane.<>
Concerns the design of cellular neural networks intended to function as associative memories. The authors consider a discrete-time version of cellular neural nets featuring simple linear thresholding neurons and the synchronous state-updating rule. The Hebbian rule is adopted as the memory design rule. Important issues, such as the memory capacity and the size of the attracting basin, are discussed. The validity of the method is illustrated by a simple example
A unified description of neural algorithms by means of general objective functions is shown to be the key to economic design of software and hardware. The compute-intensive algorithmic strings present in the dynamical equations corresponding to an objective function are to be executed by dedicated VLSI circuits. Cellular neural networks are recovered as a special case, and a corresponding general learning rule is derived
This paper addresses the problem of testing an ACNN by postulating the need for including some extra hardware to render feasible a postfabrication test. the work presented here deals with a test methodology based on adding some extra circuitry to every cell of a regular ACNN. This methodology is just an initial proposal for taking advantage of the network regularity to perform a global test which can be externally interpreted and hence has potential application for reconfiguring the network.
Presents a new kind of hierarchical binary network which may be considered as a specific case of cellular networks. Such networks are defined by primitives or processes which are personalised for network implementation in specific applications. The way in which a network may be described is provided by a process (network process) which may be used both for simulation software and for deducing the general hardware architecture of the network. Following a description of the algorithms of different simulation primitives, a network example for noise elimination in images is included to illustrate the potential of this kind of network
The stability analyses of the CNN-circuit has been carried out in the case when the sigmoid has offset, nonconstant saturation level and nonlinearity. Also the effect of the internal parasitic poles on the circuit performance has been studied through several simulations. The analyses have been verified with the SPICE simulations done on the experimental 4-by-4 CNN circuit
The application of the neural computing network concept to the minimax optimization is presented. According to the method the minimax programming problem is first transformed to the standard single-objective optimization problem and then solved by transforming it to the set of ordinary differential equations. The clustered connection-type interpretation of the neural-based minimax approach is given. The numerical results of some chosen examples are also presented. >
Explores the design of cellular neural networks (CNN) by using sampled-data analog current-mode techniques which neither requires capacitors nor resistors but just MOS transistors. The feature makes the proposed technique well suited for implementation in conventional VLSI MOS technologies. A set of building blocks is presented and their performance validated by device-level simulation results. Also, guidelines are given concerning the choice of the circuit parameters for optimum operation.< >
A current-mode CMOS circuit implementation of a cellular neural network is discussed. The implementation mimics classical cellular automata and has been designed for image processing. Signals are processed in current mode using simple current mirrors, inverters, and sources. Simulations for networks constructed have shown effectiveness in edge detection and noise removal
A new integrated circuit cellular neural network implementation having digitally or continuously selectable template coefficients is presented. Local logic and memory is added into each cell providing a simple dual computing structure (analog and digital). The variable-gain operational transconductance amplifier (OTA) is used as voltage controlled current sources to program the weighting factors of the template elements. A 4-by-4 CNN circuit is realized using the 2 μm analog CMOS-process. The circuit with different template configurations has been simulated with HSPIC
The hardware accelerator (HAC) boards using catalog programmable VLSI ICs represent a trade-off having higher reconfigurability and lower cost. This paper presents such a solution for a cellular neural network (CNN). The architecture of the present design (CNN-HAC) using 4 standard DSPs to calculate the transient response of a one-layer CNN containing 0.25-1.0 million analog neural cells is presented. The architecture and also the design principles are independent of the number of processors. The actual design was made in the form of a PC add-on board. The global control unit, which connects the board to the host firmware and communicates control signals to/from the local control units of the DSPs, was realized mainly with EPLDs. A special correspondence between the virtual processing elements-calculating the time discrete models of the analog neural cells-and the physical ones, established to work an architecture with an infrequent, one-directional interprocessor communication, is discussed in detail.<>
Three popular single step algorithms for solving the system of nonlinear differential equations of cellular neural networks are compared. The typical behaviour of these algorithms is described, example simulations are given and their relative advantages and disadvantages are discussed.< >
The moire effect, i.e. the intermodulation of signal frequencies, is generated by superposition of two or more periodic and/or quasi-periodic functions. It occurs as distortion in printing and displaying, and can be used for measurement (moire topography, stress and strain analysis). This paper describes the moire technique for medical spinal curvature estimation. Two approaches are presented: classical 2D sequential digital filter, and a cellular neural network
