Development of an image segmentation model based on a convolutional neural network

Abstract

This paper has considered a model of image segmentation using convolutional neural networks and studied the process efficiency based on models involving training the deep layers of convolutional neural networks. There are objective difficulties associated with determining the optimal characteristics of neural networks, so there is an issue related to retraining the neural network. Eliminating retraining by determining the optimal number of epochs only would not suffice since it does not provide high accuracy. The requirements for the set of images for training and model verification were defined. These requirements are best met by the image sets PASCAL VOC (United Kingdom) and NVIDIA-Aerial Drone (USA). It has been established that AlexNet (Canada) is a trained model and could perform image segmentation while object recognition reliability is insufficient. Therefore, there is a need to improve the efficiency of image segmentation. It is advisable to use the AlexNet architecture to build a specialized model, which, by changing the parameters and retraining some layers, would allow for a better process of image segmentation. Five models have been trained using the following parameters: learning speed, the number of epochs, optimization algorithm, the type of learning speed change, a gamma coefficient, a pre-trained model. A convolutional neural network has been developed to improve the accuracy and efficiency of image segmentation. Optimal neural network training parameters have been determined: learning speed is 0.0001, the number of epochs is 50, a gamma coefficient is 0.1, etc. An increase in accuracy by 3 % was achieved, which makes it possible to assert the correctness of the choice of the architecture for the developed network and the selection of parameters. That allows this network to be used for practical tasks related to image segmentation, in particular for devices with limited computing resources