Neural network models for computer vision systems

- Methods for analysis, prediction and hierarchical classification of time series based on two-level assemblies of neural networks.

- The algorithm for incremental teaching ANN identification for the synthesis of a multilayer perceptron, which avoids the heuristic process of choosing the sizes of hidden layers in a two-level ANN training model with non-stationary input data and increases the overall classification accuracy due to a step-by-step concentration on poorly identifiable data. The algorithm for the formation of training samples is proposed in order to improve the accuracy of the analysis by reducing the redundancy of the processed data. The mentioned algorithm includes: removal of time series that take constant values ​​throughout; removal of outliers; handling missing values; performing data scaling; construction of an auto-associative neural network designed for non-linear reduction of the input data dimension; division of the processed dataset into training, test and validation samples.

- Technique for testing ANN for a comparative analysis of various approaches to constructing ANN for predicting the state of technical objects based on telemetry data. Formation of training examples is carried out according to the sliding window principle. When preparing the general set of telemetric information, resampling (conversion of the initial data into a form with a fixed sampling time) and scaling (bringing the data into the range [‑1, 1]) are performed. The network outputs are also scalable. The input set is divided into a general training and final test set in the 9:1 ratio. The general training sample is randomly divided into validation (15%), test (15%), and training (70%) samples. To assess the quality of trained ANN, as well as to compare different ANN architectures, the mean square of the error and the mean absolute error are used.

- Neural network models in the form of a multilayer perceptron and neocognitron for identification and classification of topological layers objects of integrated circuits and algorithms for their training via the Lipschitz criterion that provide guaranteed convergence of training and recognition stability at the level of 94.55%.

- Models of artificial neural networks based on fuzzy logic for color separation and image analysis. The computer technology that significantly increases the reliability of recognizing vegetation areas with specified color characteristics and detecting changes in the observed area at various stages of plant disease has been created on the basis of this model. The proposed technology can be used as the basis for systems of environmental monitoring, change control and mapping of vegetation in agriculture and forestry.

- The methodology for constructing membership functions of fuzzy sets corresponding to the consequent production rules of a fuzzy inference mechanism of the Mamdani type, based on the results of processing the training sample with a heuristic algorithm of possibility clustering for object recognition in images of agricultural vegetation fields. The color, brightness of multispectral channels, leaf surface index, improved vegetation index, normalized relative biomass index, local textural features are used as the feature space. The methodology makes it possible to reduce the number of fuzzy products that form the rule base of a fuzzy inference system in comparison with the traditional method of designing fuzzy inference systems.

- The technique for recognizing objects in aerial photographs of agricultural vegetation for monitoring diseases. This technique includes the selection of the most informative features and properties of classified objects based on the interval-valued representation of the initial data about objects, the construction of object descriptors based on fuzzy logic, clustering by one of the fuzzy and possible clustering algorithms and the construction of a fuzzy inference system of the Mamdani type, the use of an inference system for classification of pixels extracted from the input images. Along with this the recognition of objects in aerial photographs of agricultural vegetation itself is performed in a fuzzy inference system. When monitoring vegetation diseases, a sequence of inference systems that corresponds to a change in the vegetation state  is generated via this technique. Experimental software tools for recognition are implemented in the Python language. The built-in tools of the OpenCV library are used for visualization. Recognition is implemented using convolutional neural networks built on the basis of the four-class neural network models SegNet and U-Net. The presented software made it possible to perform segmentation of aerial photographs with an average accuracy of 92-93%. The results can find practical use in systems of Earth remote sensing data processing for solving problems of precision farming, sustainable forest management and monitoring the state of vegetation.

- The methodology for determining the class of cluster procedures, which includes: the choice of a fuzzy clustering algorithm based on a priori assumptions about the type of cluster partition; determination of the quality indicators of the partition; determining the method for initializing the initial values ​​of the partition; evaluation using the selected indicator of a number of partitions for various parameters of the selected algorithm and highlighting the best parameters for subsequent image processing. This methodology makes it possible to determine the fuzzy clustering algorithm based on a specific thematic problem and in accordance with the characteristics of the processed data.

- The model of image processing in the form of sequential synthesis of high-resolution color images based on high-resolution panchromatic imagery and low-resolution multispectral images, calculating a basic set of features using a direct relational heuristic algorithm for feasible clustering, identification and classification based on a weighted ensemble of multilayer neural networks with one a hidden layer with a non-linear hyperbolic tangent activation function and an output layer with a soft maximum activation function. Such a model makes it possible to increase the accuracy in comparison with a single neural network and to work with a compressed indicative description of areal image objects. The experimental software is implemented in the Python language and is intended for organizing an experimental verification of the proposed synthesis technique.

- Experimental software for detecting areas of the earth's surface affected by natural disasters based on the use of SegNet and U-Net neural network models, which made it possible to achieve an accuracy of 90-93% (for U-Net). This software was implemented using the Keras library (the Tensorflow library was used for computing).

- The hierarchical neural network model for identifying the states of onboard subsystems based on ensembles of the same type of neural networks with incremental learning.

- The neural network method for searching and classifying defects in the topology of printed circuit boards, based on the use of a fuzzy description of objects and the proposed algorithm for its formation.

- The algorithm for training an ensemble of neural networks for the classification of topology defects on images of printed circuit boards and technological layers of integrated circuits. The algorithm is based on a two-level training model, which allows for training ensembles of heterogeneous neural networks. A distinctive feature of the algorithm is that not only color features in the form of separate brightness values ​​of the RGB and HSV channels are used as input data, but also reduced normalized histograms that characterize the vicinity of a certain point in the image. As a result, the use of the proposed algorithm makes it possible to minimize classification errors.

[«Image processing in the design and production of IS.pdf» (monograph)]