Методы обработки изображений микрообъектов на основе нейронных сетей и базы знаний с матрицами импликаций
Аннотация
Сформулирована проблема совершенствования и развития методологии идентификации и оптимальной обработки изображений микрообъектов. Предложены подходы к синтезу статистических, динамических, нейросетевых моделей, а также принципы адаптивного обучения рекуррентной нейронной сети на основе БЗ с матрицами импликаций и настройки. Решены задачи оптимальной обработки данных путем редуцирования избыточных связей в переменных, сжатия и использования статистических параметров, динамических, специфических характеристик изображений объектов, а также особенностей НС. Построены модифицированные вычислительные схемы регулирования параметров компонентов и алгоритмов обучения НС на основе использования механизмов настройки переменных и формирования оптимального набора обучающих данных. Разработан программный комплекс для распознавания и классификации изображений микроорганизмов в составе крови, а также проведен анализ результатов экспериментальных исследований.
Микрообъект тасвирларини идентификация килиш ва макбул ишлов бериш услубиятини такомиллаштириш ва тараккий килиш муаммоси талкин этилган. Статистик, динамик, нейротармок моделлар синтезига ёндашув, халда импликацияли матрица ва созлаш коидаларига эга билимлар базаси асосида рекуррент нейрон тармоFини (НТ) адаптив ургатиш принциплари таклиф килинган. Узгарувчиларда ортик богаашларни редукциялаш, кисиш ва объект тасвирлари статистик параметрлари, динамик, хусусий таснифлари камда НТ хусусиятлари буйича маълумотларга макбул ишлов бериш масалалари ечилган. Созлаш ва маълумотларни ургатувчи танламаларини макбул шакллантириш, НТ компонентлари камда ургатиш алгоритмларининг узгарувчиларни созлаш механизмлари асосида такомиллаштирилган кисоб схемалари курилган. Кон таркибидаги микроорганизмлар тасвирларини таниш ва синфлаштириш учун дастурий мажмуа яратилган камда тажрибавий тадкикот натижалари таклили килинган.
A topical research was carried out with aim to develop the methods for identifying and processing the data on the basis of fundamentally new approaches to solving the problems of microobjects images visualization, recognition and classification. Relevance of the research related with modeling of non-stationary objects, characterized by a large number of variables, strong variations of random time series (RTS), presence of redundancy, little informative elements is substantiated by problems solving in conditions of limited a priori information, parametric uncertainty, inaccuracy of data that are not taken into account in traditional methods of processing. The effectiveness of data processing tools is enhanced by use the mechanisms for extracting statistical parameters, dynamic properties of information, specific characteristics, hidden patterns, relationships between variables. Databases (DB) and knowledge bases (KB), formed for identification of microobjects, provide a wide opportunity for constructing the variables adaption mechanisms and interesting applications. The designed computing schemes of NN are endowed with new methods which able to cope inaccuracy of data processing without loss of performance, conform to real examples, and also adapt well to the conditions of using typical algorithms, applied software packages intended for intelligent computing. The proposed approaches are aimed to saving the user from using many iterative algorithms with time-consuming calculations caused by a large number of variables, complex structure of models for identifying a dynamic object. Methodical bases are developed.for improving the methods and algorithms for identifying of microobject image, the computational scheme of NN learning on the basis of using a recurrent network, as well as the KB of implication matrices for variables adjusting during solve the problems of microorganisms recognition and classification. A three-layer NN model is recommended for practical implementation, and it is synthesized with adequate model for microobject images identifying, adaptive network learning algorithm and algorithm of optimization based on conjugate gradient method. The generalized algorithm which combines possibilities of cubic splines is implemented for image identification and programmed algorithmic complex is realized for recognition and classification of unicellular microorganisms in the blood. The complex operates on the basis of created application software. The complex architecture is open and represents the hierarchical structure of connection, and is also developed using Builder C ++ tools.
Микрообъект тасвирларини идентификация килиш ва макбул ишлов бериш услубиятини такомиллаштириш ва тараккий килиш муаммоси талкин этилган. Статистик, динамик, нейротармок моделлар синтезига ёндашув, халда импликацияли матрица ва созлаш коидаларига эга билимлар базаси асосида рекуррент нейрон тармоFини (НТ) адаптив ургатиш принциплари таклиф килинган. Узгарувчиларда ортик богаашларни редукциялаш, кисиш ва объект тасвирлари статистик параметрлари, динамик, хусусий таснифлари камда НТ хусусиятлари буйича маълумотларга макбул ишлов бериш масалалари ечилган. Созлаш ва маълумотларни ургатувчи танламаларини макбул шакллантириш, НТ компонентлари камда ургатиш алгоритмларининг узгарувчиларни созлаш механизмлари асосида такомиллаштирилган кисоб схемалари курилган. Кон таркибидаги микроорганизмлар тасвирларини таниш ва синфлаштириш учун дастурий мажмуа яратилган камда тажрибавий тадкикот натижалари таклили килинган.
A topical research was carried out with aim to develop the methods for identifying and processing the data on the basis of fundamentally new approaches to solving the problems of microobjects images visualization, recognition and classification. Relevance of the research related with modeling of non-stationary objects, characterized by a large number of variables, strong variations of random time series (RTS), presence of redundancy, little informative elements is substantiated by problems solving in conditions of limited a priori information, parametric uncertainty, inaccuracy of data that are not taken into account in traditional methods of processing. The effectiveness of data processing tools is enhanced by use the mechanisms for extracting statistical parameters, dynamic properties of information, specific characteristics, hidden patterns, relationships between variables. Databases (DB) and knowledge bases (KB), formed for identification of microobjects, provide a wide opportunity for constructing the variables adaption mechanisms and interesting applications. The designed computing schemes of NN are endowed with new methods which able to cope inaccuracy of data processing without loss of performance, conform to real examples, and also adapt well to the conditions of using typical algorithms, applied software packages intended for intelligent computing. The proposed approaches are aimed to saving the user from using many iterative algorithms with time-consuming calculations caused by a large number of variables, complex structure of models for identifying a dynamic object. Methodical bases are developed.for improving the methods and algorithms for identifying of microobject image, the computational scheme of NN learning on the basis of using a recurrent network, as well as the KB of implication matrices for variables adjusting during solve the problems of microorganisms recognition and classification. A three-layer NN model is recommended for practical implementation, and it is synthesized with adequate model for microobject images identifying, adaptive network learning algorithm and algorithm of optimization based on conjugate gradient method. The generalized algorithm which combines possibilities of cubic splines is implemented for image identification and programmed algorithmic complex is realized for recognition and classification of unicellular microorganisms in the blood. The complex operates on the basis of created application software. The complex architecture is open and represents the hierarchical structure of connection, and is also developed using Builder C ++ tools.