Распознавание образов как реализация определенного подкласса процессов мышления

Abstract

Представлен обзор исследований проблемы распознавания образов, выполненных в течение последнего двадцатилетия в Международном научно-учебном центре информационных технологий и систем с момента его основания. Это – исследования на стыке классической проблемы распознавания и проблемы совместимости системы ограничений, известной как Constraint Satisfaction Problem. Система понятий, задач и алгоритмов на стыке этих направлений формализует определенный тип мыслительных процессов, осознанно или неосознанно выполняемых человеком или другими живыми существами.

Подано огляд досліджень проблеми розпізнавання образів, виконаних протягом останнього двадцятиріччя у Міжнародному науково-навчальному центрі інформаційних технологій та систем з дня його заснування. Це – дослідження на перетині класичної проблеми розпізнавання і проблеми несуперечності системи обмежень, відомої як Constraint Satisfaction Problem. Система понять, задач і алгоритмів на перетині цих напрямів формалізує певний тип процесів мислення, що свідомо чи несвідомо виконує людина або інші живі істоти.

The paper presents a review of pattern recognition research conducted by the International Research and Training Center for Information Technologies and Systems during the last 20 years since the moment of it's foundation. This research lies on the edge of classical pattern recognition and constraint satisfaction problems. The system of concepts, problems and algorithms produced by the merge of these fields formalizes a particular type of thought process performed by humans and other living beings. The application of Constraint Satisfaction Problem theory to pattern recognition problems has produced a breakthrough in such traditionally hard problems in computer vision as stereo vision and texture segmentation. At the same time, the merge of Constraint Satisfaction Problem theory and practical computer vision problems has led to expansion of mathematical theory of the former. First of all it has resulted in introduction of a quality function over the set of solutions and finding the best solution instead of an arbitrary one. The next generalization consisted in finding a given number of best solutions and not just a singe best solution. The paper describes methods of finding the best solution to the Weighted (Soft) Constraint Satisfaction Problem as well as the method of finding any given number of best solutions. These methods are implemented as algorithms whose domain is the set of all possible Weighted (Soft) Constraint Satisfaction Problems, i.e. a NP-hard problem class. For any given problem from the domain the algorithms either find it's solution or reject the problem. It is essential that the algorithms automatically distinguish the subdomains of their competence, i.e. the subset of problems that they do not reject. The subdomain of competence of the algorithm that finds the best solution includes the known class of submodular minimization problems but is not restricted to it. The subdomain of competence of the algorithm that finds a given number of best solutions includes the minimization of functions with a majority polymorphism but is not restricted to it.