Image texture reconstruction based on multiagent approach

Development and implementation of a multiagent system that designed for texture restoration of image background when deleting objects are examined in the thesis. General problems of photo and video reconstruction and its application are reviewed. The analysis of existing methods for restore gaps in big data arrays is held. Solution by using one of popular method of clustering analysis: special self-organizing neural network is offered. Algorithm of learning Koho- nen map is covered step by step: preparation initial map data, best matching unit choice, shifting adjacent nodes by neighborhood function, determination of training measures, ordering and fine tuning. Adaptation and modification of the Kohonen map algorithm for image restoring is proposed. Algorithms of initialization of map, choice of content of training vectors, best matching unit algorithm, neighborhood functions are highlighted in this paper. Kohonen multimap principle and cooperative solution by using multiagent paradigm is considered. Mutliagent algorithm is expanded by presegmentation step for defining of adjacent areas. Architecture of a multiagent system is developed, behavior of individual agents and communication ways are determined. Multiagent system that allows restoring background pixels of is implemented. Results of series of experiments are presented. Experiments were carried out on the database of images to search for objects of interest (CBIR database). Assessments of efficiency parameters of multimapping are presented. Evaluation of the optimal number of maps for by multimapping is based on a percentage of the original image palette coverage. The accuracy and performance of agent-based mechanism for the subsystems of training and reconstruction are represented. Assessments of accuracy metrics using SSIM, MSE, PSNR and method by pixel-per-pixel comparison are conducted. Performance estimation for various configurations of the multi-system is conducted in order to determine the most effective ones.