The relevance of the topic is due to the rapid development of processes that require fast response to atypical changes for further adjustment, while the cost of consequences is low. The task of detecting such atypical changes is called screening, a special group of which consists of tasks based on the analysis of planar images, which arise, for example, in medical diagnostics, remote sensing of the Earth, meteorology, and more. Automation of screening allows to solve problems of increase of productivity of activity of the human operator with greater efficiency, to increase the profitability of the corresponding processes.
There is a contradiction between the wide possibilities for ensuring the reliability of the solution of the problem of automated screening by neural network methods on the one hand, and the limitations of existing methods of obtaining sets of annotated data for training deep neural networks on the other.
The aim of this work is to increase the precision of classification and segmentation of planar images in automated screening systems with the development of new neural-network-based models and methods.
To achieve the aim of this work, following tasks have to be solved:
to analyze the problems of automated screening, to show the advantages and disadvantages of the neural network approach, and to substantiate the directions of research;
develop a parametric model of the data set, in particular with partially erroneous annotations, which are typical for real automated screening tasks;
to develop a model of a neural network and methods of training and prediction for the analysis of planar images;
to develop tools that implement the developed neural network models and methods, and to test and implement theoretical results in solving problems of automated screening.
The object of research is the process of analysis of planar images in automated screening systems.
The subject of research is neural network models and methods of classification and segmentation of planar images in automated screening problems.
The following scientific results were obtained within the framework of the performed research:
for the first time a parametric formalization of the data set model with partially erroneous annotations, which are characteristic of real automated screening tasks, was proposed, which allowed developing a method of generating training, test, and validation data sets;
improved the method of data set generation based on parametric model with the generation of annotations: partially erroneous for training samples and reliable - for test, which made it possible to increase the efficiency of testing neural network segmentation and classification methods in automated screening tasks;
Improved the model of convolutional neural network by adding an additional classification decoder with a normalization layer, which made it possible to build methods of multitasking learning and prediction of neural network results, which solves segmentation and classification problems simultaneously;
improved methods of multi-task learning and prediction based on an improved model of convolutional neural networks by combining classification and segmentation and the introduction constraint in calculating the segmentation loss function, which increased the precision of segmentation and classification in automated screening problems.
The results of the dissertation are implemented: in the software product SafetyRadar of VITech Lab, software products of the company "PLANET SOUTH" and in research work of the Department of Information Systems of the Institute of Computer Systems of Odessa National Polytechnic University and in the educational process of the Department of Information Systems of the Institute of Computer Systems Odessa National Polytechnic University.
