ANALYSIS OF PROPERTIES FOR HIERARCHICAL IMAGE REPRESENTATION IN MODERN COMPUTER VISION SYSTEMS
Read the full article
Widely applied computer vision methods are considered. It is established that these methods analyze images mainly on one of levels – pixel, contour, structural, etc. Drawbacks of single-level image representations restricting invariance properties of methods based on such representations and, in particular, resulting in reduction of capability to distinguish objects of different classes while solving recognition tasks are determined. Possibility to overcome these restrictions with the help of hierarchical representations is justified. Ways of developing a synthesis theory for such optimal hierarchical image analysis systems, with minimal reduction of probability of the best higher-level hypothesis selection caused by intermediate decisions are proposed. Minimization of approximation error for posterior probability distribution for higher-level hypotheses by accounting for the only best hypotheses of lower levels is proposed to perform on the base of introducing feedback connections between levels and adaptive selection of hypotheses of all levels with maximization of their mutual probability.
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License