Ivasenko I. 3. Algorithms of locally-adaptive filtering based on robust estimates for image processing

Locally-adaptive algorithms for noise removal in images corrupted by noise with mixed distribution were developed and explored in thesis. The proposed algorithms are based on piecewise-linear model of image intensity and mathematical morphology for modeling of local image objects. Developed algorithms of robust estimation of local image properties, specifically coefficients of linear regression and local variance, are based on criteria of maximum a posteriori probability. Application of the proposed robust estimation of model parameters to optimal threshold determination allows to perform reliable binary segmentation of low-contrast and noisy image fragments. It is confirmed by experimental results on radiographic images from non-destructive testing in industry and medical imaging. The fast implementation possibility by using the approximation of different structuring regions by one region makes this approach attractive for practical implementation. An important concern of this research is also the fast implementation possibility of the robust intensity estimation because nonlinear and adaptive methods are timeconsuming procedures which usually do not satisfy real-time requirements for their implementation. The problem of object segmentation with various sizes is solved by using multi-scale image processing that imply different in size structuring elements and structuring region.