The Study of Neural Networks Effective Architectures for Patents Images Processing

Abstract

In this paper, we identified the main classes of patent images that will define the neural network for comparing the corresponding classes. Training, test, and testing samples were formed for selected classes of patent images. Existing neural network architectures for working with patent images and machine learning libraries were analyzed; the deep convolutional neural network architecture and the open libraries Keras and Theano were selected to search for relevant patent images. During the implementation of this project, the neural network was trained to recognize selected classes of patent images, and an analysis of the trained model was performed. A testing methodology was chosen, the accuracy of the neural network was evaluated depending on the size of the patent images, the number of epochs and the size of the training sample. Control over neural network training was ensured by means of test and test samples so that the neural network was not retrained and worked correctly on patent images that it did not see; given an assessment of the achievements of the testing objectives.