Dávid Bálya

Cellular multiadaptive analogic architecture: a computational framework for UAV applications

An efficient adaptive algorithm in real-time applications should make optimal use of the available computing power for reaching some specific design goals. Relying on appropriate strategies, the spatial resolution/temporal rate can be traded against computational complexity; and sensitivity traded against robustness, in an adaptive process. In this paper, we present an algorithmic framework where a spatial multigrid computing is placed within a temporal multirate structure, and at each spatial grid point, the computation is based on an adaptive multiscale approach. The algorithms utilize an analogic (analog and logic) architecture consisting of a high-resolution optical sensor, a low-resolution cellular sensor-processor and a digital signal processor. The proposed framework makes the acquisition of a spatio-temporally consistent image flow possible even in case of extreme variations (relative motion) in the environment. It ideally supports the handling of various difficult problems on a moving platform including terrain identification, navigation parameter estimation, and multitarget tracking. The proposed spatio-temporal adaptation relies on a feature-based optical-flow estimation that can be efficiently calculated on available cellular nonlinear network (CNN) chips. The quality of the adaptation is evaluated compared to nonadaptive spatio-temporal behavior where the input flow is oversampled, thus resulting in redundant data processing with an unnecessary waste of computing power. We also use a visual navigation example recovering the yaw-pitch-roll parameters from motion-field estimates in order to analyze the adaptive hierarchical algorithmic framework proposed and highlight the application potentials in the area of unmanned air vehicles.

System aspects of a bionic eyeglass

In spite of the impressive advances related to retinal prostheses, there is no imminent promise to make them soon available with a realistic performance to help navigating blind persons. In our new project, we are designing a Bionic Eyeglass that is providing a wearable TeraOps visual computing power to advice visually impaired people in their daily life. In this paper the system aspects are explained. There are three different types of situations (home, office, street) and a few standard image flows (with some auditory information). The basic tasks are indoor and outdoor events, defined by blind people. Two types of cellular wave computing algorithms are used: general purpose spatial-temporal event detection by analogic subroutines developed so far, and recently developed multi-channel mammalian retinal model followed by a classifier. Typical indoor and outdoor event detection processes are being considered

Face and Eye Detection by CNN Algorithms

A novel approach to critical parts of face detection problems is given, based on analogic cellular neural network (CNN) algorithms. The proposed CNN algorithms find and help to normalize human faces effectively while their time requirement is a fraction of the previously used methods. The algorithm starts with the detection of heads on color pictures using deviations in color and structure of the human face and that of the background. By normalizing the distance and position of the reference points, all faces should be transformed into the same size and position. For normalization, eyes serve as points of reference. Other CNN algorithm finds the eyes on any grayscale image by searching characteristic features of the eyes and eye sockets. Tests made on a standard database show that the algorithm works very fast and it is reliable.

A qualitative model-framework for spatio-temporal effects in vertebrate retinas

Retinal models based on the cellular neural network (CNN) paradigm have been widely used. These neuromorphic models are based on retinal anatomy and physiology. In this paper a framework is proposed for qualitative spatio-temporal studies in vertebrate retinas, the underlying retinal anatomy is followed as closely as possible, the characteristics of the physiological models, however, are kept simple. The goal is to model the qualitative effects, since the developed models are simple, compared to a fully neuromorphic one, we have a good chance to implement them on CNN Universal Machine chips using multi-layer technology.

Feature guided visual attention with topographic array processing and neural network-based classification

Biological systems are constantly engulfed in sensory input that must be processed. Attention has evolved to cut down on the magnitude of the input and enable the agent to analyze the most important parts of the information. This is especially true for the visual system where the appropriate field of view and scale must be determined. Our system receives a video flow with considerably higher resolution than the resolution of the cellular neural net based visual microprocessor that computes the topographic features of the input. This process requires a dynamic positioning of the processing window in the video flow. We have developed a fast attention and selection algorithm that allows the system to choose the field of view and scale (zoom) level for the next frame based on the features computed from the current frame and the output of the ART or NNC-based classifiers. The algorithmic framework and hardware architecture of the system are presented along with experimental chip results for several video flows recorded in flying vehicles.

Retina model with real time implementation

This paper describes a novel retina model and its implementation. Our modeling approach is neuromorphic. The primary motivation is to deduce an algorithmic skeleton from the measurements and to make possible the on-line parameter changing. The model is tuned for half a dozen different channels with less than half percent spatial-temporal error. An of-the-shelf stand-alone system, the Bi-I, is used to real-time implement this mammalian retina model. The system is capable of computing four retina channels in video real time. The easy on-line control is solved with faders and buttons.

CNN universal machine as classificaton platform: an art-like clustering algorithm.

Fast and robust classification of feature vectors is a crucial task in a number of real-time systems. A cellular neural/nonlinear network universal machine (CNN-UM) can be very efficient as a feature detector. The next step is to post-process the results for object recognition. This paper shows how a robust classification scheme based on adaptive resonance theory (ART) can be mapped to the CNN-UM. Moreover, this mapping is general enough to include different types of feed-forward neural networks. The designed analogic CNN algorithm is capable of classifying the extracted feature vectors keeping the advantages of the ART networks, such as robust, plastic and fault-tolerant behaviors. An analogic algorithm is presented for unsupervised classification with tunable sensitivity and automatic new class creation. The algorithm is extended for supervised classification. The presented binary feature vector classification is implemented on the existing standard CNN-UM chips for fast classification. The experimental evaluation shows promising performance after 100% accuracy on the training set.

Regular small-world cellular neural networks: key properties and experiments

Small world networks have a peculiar property: even though almost all of their nodes are only locally connected, the average path length between two nodes is nearly as low as that of random networks. Networks of this type are ubiquitous in natural systems and frequently show interesting dynamics. We investigated the behavior of a new class of cellular nonlinear networks (CNNs) where all cells are locally connected to their nearest neighbors (as in a conventional CNN), but some of the nodes have long-range shortcuts as well. We give lower bounds on the number of extra links that must be added for a network to be considered as small-world and suggest an optimal topology for the added links. We also show some interesting behaviors of this network in diffusion and wave propagation phenomena.

Detecting global spatial-temporal events: saccadic suppression

You may think that while are watching a schema, your eyes move smoothly and continuously to investigate the circuitry, but that is not the case. Your eyes fixate on a part of the circuit hold for a moment then suddenly jump to a new position where a particular visual detail attracts your eyes. During this jump or saccade the eyes move rapidly and we hardly notice that, relative to our eye position, the entire visual world globally shifts. Vision is in some sense turned off during saccades, which is called saccadic suppression. This paper presents a neuromorphic algorithm to detect these spatial-temporal events from the input flow. The algorithm of the saccadic suppression event detection is implemented on the standard cellular nonlinear network universal machine environment. The processing time is roughly one millisecond on the ACE4k CNN-UM chip in the Aladdin environment, and therefore it can serve as a common function in any real-time spatial-temporal algorithm. The saccadic suppression event detector can be used in different areas such as video shot detection or to define the instant when the time-dependent part of the algorithm, e.g. homotopy, has to be initialized.

Neuromorphic algorithm for retinal implants

The paper focuses on the input transformation for retinal implants. It describes a novel retina model, based on neurobiological measurements. Our modeling approach is neuromorphic, however the primary motivation is to derive an algorithmic skeleton from the animal retina measurements. The on-line, real-time implementation of the algorithm is given for an off-the-shelf stand-alone system, called Bi-I