Universal methodology for optimizing the system of recording, processing and machine analysis of graphic data

Abstract

The analysis of modern graphic data processing systems was carried out and the methodological bases for optimizing the visual object recognition procedure were determined. The increase in the efficiency of machine analysis of graphic data using neural network algorithms under the conditions of recovery procedures, setting parameters and preliminary segmentation of the image matrix is noted. A comprehensive technique for automating the processing of graphic data was proposed, which included the stages of image pre-processing based on morphological methods, mathematical models and neural network algorithms, as well as the formation of a training sample for learning neural network algorithms, the selection of visual objects and the definition of neural network architecture and the scheme of learning neural network algorithms. At the same time, it was proposed to restore the image matrix through the construction of a mathematical model of noise distribution and the use of spatial filtering methods, morphological methods of erosion and dilation, which are adapted to work with color images through the inclusion in the algorithm of the method of connective components and threshold methods, as well as neural network architecture autoencoder type. In order to determine the optimal parameters for setting up the neural network architecture, the appropriate mathematical apparatus was formalized at the level of defining the sets of variables characterizing the neurons of the layers of the neural network, the weighting coefficients and the activation function. Similarly, the formalization of the procedure for preparing the training and test sample of learning the neural network algorithm included the determination of the type of objects of the training sample according to the format of their presentation, as well as their preliminary classification. Thus, the developed scheme, based on the definition of target indicators, can be used to evaluate the effectiveness of software and neural network algorithms used for a wide class of tasks on the selection of visual objects. The optimization of the general system of machine analysis is formalized as the task of determining the global extrema of the target functions by changing the parameters characterizing the operation of software and neural network algorithms.