Skip to main content

Nonlinear and Adaptive Techniques in Digital Image Processing, Analysis and Computer Vision

Objective

The application of nonlinear techniques in image sequence processing, color image processing as well as in black and white (BW) image processing is studied. Nonlinear digital filter adaptation is also examined. Novel adaptation techniques are investigated and novel adaptive nonlinear structures are derived based on the merging of nonlinear filter families and neural networks. Another aim is the enhancement of research efforts of the participating groups though cooperation and the organisation of specialised workshops and short courses.
Nonlinear and adaptive techniques are being developed in digital image processing, digital image analysis and computer vision. The research encompasses polynomial filters, order statistics filters, mathematical morphology, invariant transforms and neural networks.

Interaction and exchange of research results have been obtained between nonlinear filter classes (notably neural networks and order statistics filter). The neural filters have emerged as a combination of stack filters and neural network techniques. Simulated annealing techniques have been used in the adaptation of order statistic filters. The LVQ neural nets have been combined with human visual system characteristics for image coding. They have also been combined with median filters to produce the median LVQ algorithms that have enhanced robustness characteristics. LVQ segmentation has also been combined with quadratic signal adaptive filters for ultrasonic image processing. Finally, fuzzy techniques have been incorporated in digital filters to produce fuzzy filters. In the area of optimal nonlinear filter design, optimal multichannel L-filters and weighted order statistics filters (WOS) have been designed. Significant progress has been made in the adaptation of L-filters as well as of stack and WOS filters.

Polynomial (quadratic) filters have been designed for digital image enhancement before binarization. The quadratic operators were capable of preserving image details. Nonlinear image restoration has successfully been used in the restoration of Poisson processes that are found in weak image sources (astronomy and microscopy images). Nonlinear algorithms have also been developed to obtain super deep resolutions from multiframe images.
APPROACH AND METHODS

Adaptation methods for polynomial and order statistic filters have been developed (notably L-filters, WOS and stack filters). Polynomial (quadratic) filters have been developed for digital image enhancement (preprocessing for binarisation). Polynomial filters have also been applied to ultrasonic image processing. Neural network techniques have been used for filter adaptation and image compression. Merging of classes of neural and nonlinear image processing techniques has been performed. Nonlinear approaches to the increase of image resolution from multiple image frames as well as to the restoration of images produced by weak image sources have been studied with applications in microscopy and astronomy.

Further studies on the application of these techniques to color image processing as well as to image sequence processing will be performed. Motion adaptive filtering schemes will also be studied. In the case of nonlinear image analysis and computer vision, two approaches will be investigated: mathematical morphology techniques and nonlinear position invariant transforms. The link between morphology, computational geometry and shape decomposition will be studied. Parallel implementations will be studied for all nonlinear processing and analysis schemes to be developed in this project. Theoretical analysis of the parallelisation as well as simulation results on images and image sequences will be performed.

POTENTIAL

The project is expected to produce novel contributions in all the above areas and to address the need for a unifying theory for the various nonlinear filter classes by the interaction and exchange of research results in the various filter classes. Such a cross-fertilisation is expected to give valuable results. The parallelisation of algorithms for nonlinear digital image processing and analysis will contribute to the development of fast parallel unifying methods and architectures for this particular area. Although basic research will be performed, the project is expected to produce original technical contributions that will be of use to the industry in applications such as telecommunications and HDTV. The results will be disseminated through scientific publications, the organisation of a workshop and through the medium of a short course designed to promote technology transfer to European industry.

Coordinator

ARISTOTLE UNIVERSITY OF THESSALONIKI
Address
University Campus, Egnatia Street, Administration
54006 Thessaloniki
Greece

Participants (5)

FORTH RESEARCH CENTER OF CRETE
Greece
Address
, 1527
71110 Heraklion
TAMPERE UNIVERSITY OF TECHNOLOGY
Finland
Address
, 527
33101 Tampere
Technische Universität Hamburg-Harburg
Germany
Address
Harburger Schloßstraße 20
21071 Hamburg
UNIVERSITA DEGLI STUDI DI TRIESTE
Italy
Address
Piazzale Europa 1
34125 Trieste
UNIVERSITY OF STRATHCLYDE
United Kingdom
Address
16 Richmond Street
G1 IXQ Glasgow