Generic tools for digital image processing and computer vision

Project Details

Description

The research on 'Generic Tools' consists of the development of new image processing algorithms and of the (parameter) tuning of both new and existing algorithms, so that they can be applied on different types of images and in various application domains. The main classes of algorithms that are studied at ETRO/IRIS are image segmentation & labeling, neural networks, computation of optical/normal flow, Wavelet transformations, coding in the context of compression. The specific output of the project 'Generic Tools' as such, is the documentation on the behavior and the performance of the studied algorithms and on the appropriate tuning of the algorithm parameters. Except for the algorithm modules, 'Generic Tools' is also concerned with tools for the efficient implementation of algorithms (parallellization). Finally, generic technology for developing vision applications in high-level environments emerged from previous research in the context of 'Knowledge Engineering in Diagnostic Imaging' (e.g., the use of expert systems such as blackboard systems (see 9.)). 1. Pyramid segmentation: fundamental properties like shift, scale, and rotation variance, connectivity of the segments have been studied. Extension to pseudo-3D and 3D images has been analyzed. The difficult control and intrinsically unreliable results of the segmentation have been explained; 2. Segmentation seen as a statistical classification problem: Although the formalization of the problem (selection of the classification parameters either based on a model or on heuristics; definition of the objects to be classified - pixels, regions, ...) and the outline of the procedure (learning phase, test phase, application phase) are always very similar, different particularities have to be taken into account dependent on the specific problem which is investigated; 3. Transform compression and pixel classification in a feature space : several alternative transforms have been tested (e. g. Karhunen Loève Transform, Kittler and Young Transform, Fisher Transform, transforms based on parametric fitting). The application area is : compact representation and segmentation of time varying image sequences; 4. Edgmentation: a modified split and merge algorithm was developed for image enhancement, data reduction and segmentation in medical applications. It was applied on MRImages and 3D and 4D SPECT images of the heart for the calculation of volume-time curves of the 4 heart cavities; 5. Cavity Detector (CD): CD is a generic sequence of algorithms leading to segmentation. The technique allows handling of bad representation of walls (e. g. due to movement artifacts in cardiac MRI) or objects which are not completely surrounded by a wall (e.g. the eye cavities in brain CT-scans). In almost all the cases the automated segmentation is superior to the purely manual one; 6. Labeling schemes: Image segments are classified and merged, based on Probability Distribution Maps (PDM) which store the model information acquired during a learning phase; 7. Artificial neural network based local edge detectors for cavity delineation in cardiac MRImages; 8. Evaluation of different optical flow algorithms (reliability and accuracy analysis); 9. Knowledge engineering in diagnostic imaging : this project essentially aimed at the design of a general framework for the development of expert systems in the diagnostic imaging domain. It involved the elaboration of a uniform representation for the structure and semantics of data objects and Image Analysis (IA) algorithms, allowing their symbolic manipulation by a knowledge based reasoning system. This system also had to deal with uncertainty and had to be able to exploit complementarities, redundancies and contradictions of the output of the various Image Analysis (IA) algorithms through cooperative strategies. A flexible scheme for the object representation ---the Feature Dictionary --- has been elaborated, allowing the dynamic storage and update of features per image object, and accessible by an automatic feature conversion routine, which facilitates the integration of many different but equivalent feature-data-representations. Different aspects of the design of a reasoning system for medical image analysis have been studied including user interfacing, automated data format transformation and data fusion, based upon models like the Bayesian, Dempster-Shafer, Multi-valued logic, and the scoring model. An expert system, using a blackboard (BB) architecture has been developed and applied successfully to the case study of the segmentation and labeling of anatomically defined objects in brain CT scans (Ph.D. thesis of Li Hongyi).
AcronymOZR646
StatusFinished
Effective start/end date1/01/9631/12/97

Flemish discipline codes

  • Mathematical sciences
  • Electrical and electronic engineering

Keywords

  • wavelet transformations
  • image coding
  • neural networks
  • image segmentation