| Abstract [eng] |
The purpose of this project was to apply the classification algorithm to determine the type and country of identity documents from visual material, using high classification accuracy methods. Convolutional Neural Networks have been used to implement document classification, which allows achieving the highest results for solving image classification problems in comparison with other algorithms such as random forests, decision trees, Support Vector Machines. The analyzes were performed to find out the current situation, used technologies, machine learning algorithms, and detailed convolutional neural networks. The relevance of the problem is presented by analyzing the causes and assessing the need for the system. The task was solved using CNN modified AlexNet architecture by changing the number of layers and hyperparameters. The speed and accuracy of the system are evaluated and the research results are presented. Other convolutional network architectures, such as VGG16, VGG19, ResNet50, ResNet152, Inception V3, are also used in the experimental study. To validate the accuracy of the classification, the model was validated and tested using cross-validation. Experiments were performed to evaluate the performance of the models using poor-quality photos. The characteristics of different architectural models are compared: the size of the trained model, the time of testing and training, the number of parameters, and the accuracy. After analyzing the research results, the most appropriate model for solving the task of classifying identity documents using visual material was determined. |
