Daria Ershova

Document image binarization task requires neural networks that can effectively analyze large image areas to separate text from complex backgrounds. The traditional approach of expanding the receptive field by increasing the depth of the network results in a significant increase in computational complexity and model size. The latter is most critical for mobile devices and web applications, where speed and size are of paramount importance. To address this problem, we propose IIR-BinNet, a novel ultra-lightweight neural network architecture that leverages multidirectional learnable Infinite Impulse Response (IIR) filters. IIR layers allow efficient multi-scale feature extraction based on different receptive fields provided by each filter within the same IIR layer. The proposed approach was tested in the task of binarization of complex images of historical documents, which requires understanding of the global context. With only 49K parameters, our model is at least 40 times smaller than U-Net-based solutions while maintaining competitive performance on the DIBCO 2017 and H-DIBCO 2018 benchmarks. The results show that IIR-BinNet effectively replaces deep convolutional stacks with a block of learnable multidirectional IIR filters, enabling competitive performance among lightweight approaches on DIBCO 2017 and stronger results on H-DIBCO 2018 within the same class of models. On the H-DIBCO 2018, IIR-BinNet achieves an F-measure of 90.37%, compared to 89.36% for the baseline U-Net. Among lightweight specialized models evaluated on H-DIBCO 2018, it achieves the best reported performance with a 35% reduction in DRD score.

Existing approaches to barcode detection have a number of disadvantages, including being tied to specific types of barcodes, computational complexity or low detection accuracy. In this paper, we propose YOLO-Barcode – a deep learning model inspired by the You Only Look Once approach that allows to achieve high detection accuracy with real-time performance on mobile devices. The proposed model copes well with a large number of densely spaced barcodes, as well as highly elongated one-dimensional barcodes. YOLO-Barcode not only successfully detects the huge variety of barcode types, but also classifies them. Comparing with the previous universal barcode detector DilatedModel based on semantic segmentation, the YOLO-Barcode is 4 times faster and achieves state-of-the-art accuracy on the ZVZ-real public dataset: 98.6% versus 88.9% by F1-score. The analysis of existing publicly available datasets reveals the absence of many real-life scenarios of mobile barcode reading. To fill this gap, the new “SE-barcode” dataset is presented. The proposed model, used as a baseline, achieves a 92.11% by F1-score on this dataset.

The search for stamps on images is necessary to verify the authenticity of a document and extract valuable textual information contained in them. Despite the vast number of methods for detecting stamps, most of them are not universal and are limited either by the characteristics of the stamp or by the computational complexity of the approach. The last criterion is extremely relevant, since document recognition is actively transferred from server solutions to mobile devices with low computing power and limited battery life. In view of this, in this paper we propose a compact and computationally efficient neural network model for searching for stamps of any shape and color. Based on the popular neural network detector YOLO, our model contains only 128 thousand parameters. We demonstrate the effectiveness of the model in an experiment on the public dataset SPODS (scanned pseudo-official dataset), containing high-resolution scans of documents with stamps.

Various government and commercial services, including, but not limited to, e-government, fintech, banking, and sharing economy services, widely use smartphones to simplify service access and user authorization. Many organizations involved in these areas use identity document analysis systems in order to improve user personal-data-input processes. The tasks of such systems are not only ID document data recognition and extraction but also fraud prevention by detecting document forgery or by checking whether the document is genuine. Modern systems of this kind are often expected to operate in unconstrained environments. A significant amount of research has been published on the topic of mobile ID document analysis, but the main difficulty for such research is the lack of public datasets due to the fact that the subject is protected by security requirements. In this paper, we present the DLC-2021 dataset, which consists of 1424 video clips captured in a wide range of real-world conditions, focused on tasks relating to ID document forensics. The novelty of the dataset is that it contains shots from video with color laminated mock ID documents, color unlaminated copies, grayscale unlaminated copies, and screen recaptures of the documents. The proposed dataset complies with the GDPR because it contains images of synthetic IDs with generated owner photos and artificial personal information. For the presented dataset, benchmark baselines are provided for tasks such as screen recapture detection and glare detection. The data presented are openly available in Zenodo.