Krzysztof Slot

Determination of cellular neural networks parameters for feature detection of two-dimensional images

Issues involved in cellular neural net design are discussed, and recommendations are made for parameter choice. Inherent errors of detection are pointed out and a method for their reduction is proposed. Complex signal processing in the net from the point of view of error occurrences is also discussed. A simple cell architecture is introduced, and its modification, appropriate for complex signal processing, is presented.<<ETX>>

Keypoints derivation for object class detection with SIFT algorithm

The following paper proposes a procedure for SIFT keypoints derivation for the purpose of object class detection. The main idea of the method is to build appropriate object class keypoints by extracting information that corresponds to characteristic class features. The proposed procedure is composed of two main steps: clustering of similar SIFT keypoints and derivation of appropriate keypoint descriptors. Face detection in images has been selected as a sample application for the proposed approach performance evaluation.

Optically realized feedback cellular neural networks

An optical implementation of cellular neural networks (CNNs), defined by a template with a nonzero feedback operator, is discussed. Two design rules which consider a transient duration of optically realized CNNs are formulated. Possible methods of improving the processing properties of optically realized CNNs are presented. The application of the two design rules can result in an increase in the processing speed of some optically realized feedback CNN applications. A signal dynamic range shift in the electronic feedback path of the system yields an increase in the resolution of images which are subject to optical processing.<<ETX>>

Object Representation Using Geodesic Levels

The following paper explores possibilities of using geodesic distance-based description of two-dimensional image objects. The main idea of the proposed representation is to create a set of equally-spaced contours composed of image points that are equidistant, in terms of geodesic distance, from the objects boundary. Due to geodesic distance properties, the proposed representation is expected to be robust against several types of nonlinear object deformations, thus being attractive for object recognition purposes.

TEXTURE GENERATION USING CELLULAR NEURAL NETWORKS

The following paper introduces an application of Cellular Neural Networks for the generation of predetermined stochastic textures. The key element for the task realization is an appropriate selection of template elements, which should provide a transformation of initial, random CNN state into a stable equilibrium, featuring desired perceptual properties. A template derivation procedure comprises two steps: linear CNN design, followed by a template-refinement procedure that involves nonlinear optimization. In addition, a procedure that extends CNN texture rendition capabilities into a realm of non-pure stochastic textures is proposed.

Computer-assisted analysis of the extracellular matrix of connective tissue

The new computerized imaging, circular polarized light microscopy technique was developed to measure the orientation of collagen fibers in images of serial sections of connective tissue. The system consists of a modified Olympus BX50 polarized microscope, a Sony AVC-D7 video camera, and a Silicon Graphics Indy computer. Both methods required the initial segmentation of fibers and used binary images. Segments of fiber midlines were traced with vertical and horizontal scanlines, or alternately the whole midlines were identified recursively from the Euclidean Distance Map of the image suing the novel definition of the Medial Axis Transform. The last technique produced connected midlines of the fibers and handled sinuous fibers well. The fiber midlines produced by this technique were traversed by a midline traversal algorithm , and the orientation distribution was obtained by least squares line fitting. The accuracy of the developed techniques was evaluated against synthetic images, composed of straight lines and sinuous curves. Kupiers statistic was used to evaluate the consistency of the fiber orientation calculations. Statistical analysis of the results showed, that the proposed Medial Axis Transform with Hilditchs connectivity preserving skeletonization produced the most accurate results. The developed method was used to measure collagen fiber orientation in microscopy images of canine meniscus, porcine aortic valve leaflet, bovine pericardium and bio- textiles.

Multiple-input OTA based circuit for cellular neural network-implementation in VLSI CMOS technology

An operational transconductance amplifier with multiple inputs, each of them realized by means of two MOSFETs only, suitable for implementation of cellular neural networks in CMOS technology is described. Preliminary analysis of a short-channel circuit augmented by SPICE simulation is given. Results show the potential for designing chip area effective networks applicable to various tasks of image processing and pattern recognition in real time. Additional research, in particular in the area of VLSI implementation, is needed.<<ETX>>

Application of voiced-speech variability descriptors to emotion recognition

The following paper examines a possibility of applying phone-pronunciation variability descriptors in emotion classification. The proposed group of descriptors comprises a set of statistical parameters of Poincare maps, which are derived for evolution of formant-frequencies and energy of voiced-speech segments. Poincare maps are represented by means of four different parameters that summarize various aspects of plots scatter. It has been shown that incorporation of the proposed features into a set of commonly-used emotional-speech descriptors, results in a substantial, ten-percent increase in emotion classification performance - recognition rates are at the order of 80% for six-category, speaker independent experiments.

CNN-based object recognition with deformable grids and multiple-feature image representation

The following paper presents a new direction that opens for deformable grid-based object recognition methods, due to introduction of their efficient, parallel implementations. A substantial increase in object recognition performance can be expected when several different features are used to build a class prototype. This would imply extending complexity of image analysis, through an application of several image characteristics in image-model matching. To make such an approach computationally feasible, a CNN is considered as ultra-fast tool for performing grid-matching process. Sample task of face recognition, which is well-suited for being tackled with deformable grids, is used to evaluate a performance of the proposed approach, yielding an expected increase in correct classification rate.

Lee-algorithm based path replanner for dynamic environments

The presented paper introduces a new procedure, based on well-known Lees algorithm, for vehicle collisionless path determination in dynamically-changing environments. Such routes need to be updated on-the-fly to take into account moving obstacles or newly detected objects. The main idea of the proposed approach is to first, identify regions where environment has changed and to execute a procedure of selective path updates. As a result, an optimal path can be derived at a computational expense comparable to the heuristic Lifelong A* search.