Короткий опис (реферат):
The work is devoted to the development of new integrated morphological-neural network algorithms based on nonlinear equivalence metrics for recognizing halftone black-and-white images of multi-symbol identification objects, which are widely used in automated production and various logistics transport systems for the purpose of ensuring the accounting and security of transport units or their identification objects. These algorithms are based on the use of linear (nonlinear) equivalence (non-equivalence) metrics and normalized spatially dependent similarity functions (similarity or dissimilarity) of matrix arrays or data in the form of images, as criterion (discriminant) functions. Based on the review of related works, the aspects, specifics and advantages of using such metrics and functions to solve research problems were analyzed, the necessity and relevance of which were technically justified and formulated. Two groups of algorithms for recognizing images of multi-symbol identification objects have been developed and presented, the main procedures in these algorithms have been described, their features and differences have been shown. In the article, the authors present the results of model experiments of the proposed algorithms using the created auxiliary program for studying the features of the application of algorithms, for checking the quality and accuracy characteristics of image recognition when exposed to various types and strengths of interference, for verifying the adequacy of models and metrics. A comparative analysis of the obtained modeling results has been performed taking into account various conditions and interference. A program for recognizing identification numbers on locking and sealing devices has been developed and tested in real production conditions. The operation of the program and its user interface are demonstrated by screenshots. The results obtained showed that such proposed integrated morphological-neural network algorithms based on equivalent nonlinear metrics have better discriminant properties, accuracy and performance characteristics, especially in the case of a significant (up to 40%) conditional level of interference power acting on the recognized images.
