In this paper, we introduce an “on the device” text line recognition framework that is designed for mobile or embedded systems. We consider per-character segmentation as a language-independent problem and individual character recognition as a language-dependent one. Thus, the proposed solution is based on two separate artificial neural networks (ANN) and dynamic programming instead of employing image processing methods for the segmentation step or end-to-end ANN. To satisfy the tight constraints on memory size imposed by embedded systems and to avoid overfitting, we employ ANNs with a small number of trainable parameters. The primary purpose of our framework is the recognition of low-quality images of identity documents with complex backgrounds and a variety of languages and fonts. We demonstrate that our solution shows high recognition accuracy on natural datasets even being trained on purely synthetic data. We use MIDV-500 and Census 1961 Project datasets for text line recognition. The proposed method considerably surpasses the algorithmic method implemented in Tesseract 3.05, the LSTM method (Tesseract 4.00), and unpublished method used in the ABBYY FineReader 15 system. Also, our framework is faster than other compared solutions. We show the language-independence of our segmenter with the experiment with Cyrillic, Armenian, and Chinese text lines.

We propose a lightweight noise-canceling filtering neural network that implements the filtering stage in the algorithm for tomographic reconstruction of convolution and backprojection (Filtered BackProjection—FBP). We substantiate the neural network architecture, selected on the basis of the possibility of approximating the ramp filtering operation with sufficient accuracy. The network performance has been demonstrated using synthetic data that mimics low-exposure tomographic projections. The quantum nature of X-ray radiation, the exposure time of one frame, and the nonlinear response of the ionizing radiation detector are taken into account when generating the synthetic data. The reconstruction time using the proposed network is 11 times shorter than that of the heavy networks selected for comparison, with the reconstruction quality in the \(SSIM\) metric above 0.9.

This work is devoted to methods for improving the quality of artificial neural networks, developed by a group of scientists led by V.L. Arlazarov. It describes various approaches both to training the networks themselves, from applying regularization to using new types of layers, and to preparing training data. This paper presents an effective method of data augmentation, as well as a neural network method for generating text data for training, which allows taking into account the properties of the language model in text recognition tasks. The applicability of metric networks to solving the classification problem and methods for generating and balancing training data for training such networks by creating Siamese and triplet schemes using contrastive loss and triplet loss functions are described. A separate section is devoted to research on the use of the fast Hough transform in neural networks as one of the methods for extracting nontrivial features. The paper also describes studies of a block convolutional layer, a method for constructing estimates of the presence of cuts between characters in the text segmentation problem, and a method for increasing the generalization ability of networks using orthogonalization.

In this paper, we deal with the document machine readable zone (MRZ) detection in the images taken by smartphones. These images could have low quality and projective distortions. Current solutions for the considered task are mostly image processing algorithms. All of them have precision problems in various use cases. Known neural network methods could address this issue, but impose strict requirements for computation power which is critical for on-device inference. Therefore, we propose a method that combines neural network and image processing that is both lightweight and accurate. The network is trained to process the input images and obtain the heatmap of MRZ characters. After that, we use connected components analysis to merge the characters into lines and evaluate the MRZ bounding box. The proposed neural network is a light version of the Hough encoder—architecture that was designed to work with projectively distorted images. Our network is 1.7 times smaller than the original Hough encoder and more than 100 times smaller in comparison with typical autoencoders, which makes it possible to use the proposed method in embedded devices. Experiments were held on the open synthetic dataset of the documents with 3 types of MRZ. Our results show that our method gives high quality on the test dataset, and have less trainable parameter compared to the common solution: Unet.