Vlasenko N. Feature descriptions models and their transformation in image recognition

The dissertation work is devoted to the development of efficient methods for construction, analysis and transformation of feature descriptions of images to improve recognition performance. A method for construction of compact informative feature-based descriptions based on models of data representation and processing in Walsh orthogonal basis was introduced, which makes it possible to recognize images of objects with high performance without reducing of noise immunity level. An approach of feature descriptions compression using the stability criterion, which provides a significant gain in the overall processing time and reduce the amount of descriptions preserving the required level of correct classification was proposed. Improved models for the construction of descriptions similarities and ranking by selecting a finite tuple of the most similar components, and using the principle of nearest neighbors classification were suggested. Multi-objective optimization model for the optimal choice of threshold for description elements equivalence was improved. The implementation of this model ensures adaptation to the image database and improves the reliability of recognition. Practical problems to improve performance of the classifications of images obtained by remote sensing of natural environments of aerospace media and during an estimation of coating roads thickness due to optimal signal processing ground penetrating radar were resolved, as well as an automation and reducing time of medical plants image processing.