The Method of Semantic Image Segmentation Using Neural Networks

Currently, the means of semantic segmentation of images, which are based on the use of neural networks, are increasingly being used in computer systems for various purposes. Despite significant progress in this industry, one of the most important unsolved problems is the task of adapting a neural network model to the conditions for selecting an object mask in an image. The features of such a task necessitate determining the type and parameters of convolutional neural networks underlying the encoder and decoder. As a result of the research, an appropriate method has been developed that allows adapting the neural network encoder and decoder to the following conditions of the segmentation problem: image size, number of color channels, acceptable minimum segmentation accuracy, acceptable maximum computational complexity of segmentation, the need to label segments, the need to select several segments, the need to select deformed , displaced and rotated objects, allowable maximum computational complexity of training a neural network model, allowable training time for a neural network model. The main stages of the method are related to the following procedures: determination of the list of image parameters to be registered; formation of training example parameters for the neural network model used for object selection; determination of the type of CNN encoder and decoder that are most effective under the conditions of the given task; formation of a representative educational sample; substantiation of the parameters that should be used to assess the accuracy of selection; calculation of the values of the design parameters of the CNN of the specified type for the encoder and decoder; assessment of the accuracy of selection and, if necessary, refinement of the architecture of the neural network model. The developed method was verified experimentally on examples of semantic segmentation of images containing objects such as a car. The obtained experimental results show that the application of the proposed method allows, avoiding complex long-term experiments, to build a NN that, with a sufficiently short training period, ensures the achievement of image segmentation accuracy of about 0.8, which corresponds to the best systems of similar purpose. It is shown that it is advisable to correlate the ways of further research with the development of approaches to the use of special modules such as ResNet, Inception and mechanisms of the Partial convolution type used in modern types of deep neural networks to increase their computational efficiency in the encoder and decoder.