Abstract:

This paper considers a model of object recognition in images using convolutional neural networks; the efficiency of the model-based process involving the training of deep layers in convolutional neural networks has been studied. There are objective difficulties associated with determining the optimal characteristics of neural networks, so there is an issue related to retraining a neural network. Eliminating the retraining by determining only the optimal number of epochs is insufficient since it does not provide high accuracy. The requirements for the set of images for model training and verification have been defined. These requirements are better met by the INRIA image set (France). GoogLeNet (USA) has been established to be a trained model that can perform object recognition on images but the object recognition reliability is insufficient. Therefore, it becomes necessary to improve the effectiveness of object recognition in images. It is advisable to use the GoogLeNet architecture to build a specialized model that, by changing the parameters and retraining some layers, could allow for better recognition of objects in images. Ten models were trained using the following parameters: learning speed, the number of epochs, an optimization algorithm, the type of learning speed change, a gamma or power coefficient, a pre-trained model. A convolutional neural network has been developed to improve the precision and efficiency of object recognition in images. The optimal neural network training parameters were determined: training speed, 0.000025; the number of epochs, 100; a power coefficient, 0.25, etc. A 3 % increase in precision was obtained, which makes it possible to assert the proper choice of the architecture for the developed network and the selection of its parameters. That allows this network to be used for practical tasks of object recognition in images. Author Biographies Bogdan Knysh, Vinnytsia National Technical University PhD, Associate Professor Department of Electronics and Nanosystems

Keywords: