Deep Learning for Automatic Classification of Identification Documents

This paper presents a general approach for identification documents classification using deep learning models. Our study gives an explanation of the main steps that need to be followed in order to implement a classification deep learning model. We have used convolution neural networks to extract features from raw image pixels on private datasets of identification documents. The implemented models use different techniques to preprocess the images in order to improve the classification performance on the test dataset and also techniques that can offer a better generalization of the models on the classification task. The experiments demonstrate that the training time-efficiency and accuracy of the models depends on the size, numbers of the pattern for each category and type of the image preprocessing. Various techniques of optimization have been applied to improve the model’s performance and as a result we achieved the best classification accuracy of 90.4% on the test dataset.