Макаров Илья Андреевич

Human-in-the-Loop Egyptology: A System for Ancient Egyptian Text Study

2026 · CHAPTER · en

Мы представляем прототип веб-системы с участием человека для изучения древнеегипетских иероглифических текстов, которая интегрирует конвейер обработки изображений в текст в интерактивное рабочее пространство для итеративной доработки. В ходе предварительного исследования с участием египтологов и студентов система позволила ускорить работу и получить более качественные результаты по сравнению с ручным процессом.

Advancing Sequential Manga Colorization for AR Through Data Synthesis

2025 · ARTICLE · en

Manga colorization in augmented reality (AR) environments presents unique challenges, particularly when colorizing manga pages captured in photos under various real-world conditions. Testing models in AR settings for manga colorization has been a significant challenge, primarily because of the absence of suitable datasets tailored for this task. To address this, we propose a benchmark for evaluating existing colorization models. We first collected a relatively small dataset of manga book photos taken in settings suitable for AR applications. Then, we developed a method that leverages a pretrained diffusion model to generate synthetic photos from scans of manga pages. Using large datasets of manga scans, we created an extensive synthetic dataset. Combining both real and synthetic data, we established a comprehensive benchmark for manga colorization in AR scenarios. We tested existing models for natural image and manga colorization on this benchmark. As a result, our evaluation showed that current models are not well-suited for AR-based colorization tasks, indicating a need for further improvement.

Closing the Domain Gap in Manga Colorization via Aligned Paired Dataset

2025 · CHAPTER · en

This paper addresses the challenge of artwork colorization by proposing a benchmark for manga colorization using real black-and-white and colorized image pairs. Color images are widely recognized for their ability to capture attention and improve memory retention, yet the manual process of colorization is labor-intensive. Deep learning methods for supervised image-to-image translation offer a promising solution, relying on aligned pairs of black-and-white and color images for training. However, these pairs are often generated synthetically, introducing a domain gap that limits model performance. To address this, we explore the use of real data, proposing a method for creating such datasets. Our benchmarks reveal that models trained on real data significantly outperform those trained on synthetic pairs. Furthermore, we present a pipeline for text removal and panel segmentation, streamlining the comic colorization process. These contributions aim to enhance the generalization and applicability of deep learning models for artwork colorization.

SensorDBSCAN: Semi-Supervised Active Learning Powered Method for Anomaly Detection and Diagnosis

2025 · ARTICLE · en

Fault detection and diagnosis (FDD) is a critical challenge in industrial processes aimed at minimizing risks such as safety hazards, costly downtime, and suboptimal production. Traditional supervised FDD methods offer great performance while heavily relying on large volumes of labeled data, whereas unsupervised methods do not depend on labeled data, though are inferior in performance compared to supervised ones. In this paper, we propose SensorDBSCAN, a novel semi-supervised method for anomaly detection and diagnosis. The key innovation lies in achieving good performance with minimal labeled data - less than 1% of the dataset - by leveraging active and contrastive learning techniques. The proposed approach combines a transformer-based encoder trained with a triplet-based contrastive learning objective and the classical density-based clustering algorithm DBSCAN, enabling strong feature extraction, efficient and interpretable feature space organization and simple clustering algorithm. Unlike existing methods, SensorDBSCAN eliminates the need for manual labeling large amounts of data, cluster analysis, and pre-defining cluster numbers, providing greater usability in real-world cases. We validate the effectiveness of our method on the Tennessee Eastman Process (TEP) and its advanced simulations (TEP Rieth and TEP Rieker). SensorDBSCAN demonstrates better performance on well-known and realistic datasets, reducing labeling requirements while maintaining high accuracy of fault detection and diagnostics. The code is available at https://github.com/K0mp0t/sensordbscan.

Explainable Document Classification via Concept Whitening and Stable Graph Patterns

2025 · ARTICLE · en

This paper proposes a novel explainable document classification framework that integrates Concept Whitening (CW) with graph concepts that are derived from stable graph patterns, and extracted via methods based on Formal Concept Analysis (FCA) and pattern structures. Document graphs are constructed using Abstract Meaning Representation (AMR) graphs, from which graph concepts are extracted and aligned with the latent space axes of Graph Neural Networks (GNNs) using CW. We investigate four types of graph concepts for their effect on concept alignment: frequent subgraphs, graph pattern concepts, filtered equivalence classes, and closed subgraphs. A novel filtration mechanism based on support, along with a custom penalty metric, is proposed to refine graph concepts for maximizing concept alignment. Experiments on the 10 Newsgroups and BBC Sport datasets show that our document graphs effectively capture both structural and semantic information, thereby supporting competitive classification performance across multiple GNN model architectures and configurations. For the 10 Newsgroups dataset, GNN models equipped with a CW module show an average increase of 0.7599 in the macro-averaged F1 score of the Concept Alignment Performance (CAP) metric, with an average drop of only 0.0025 in the document classification macro-averaged F1 score. Similarly, on the BBC Sport dataset, the average CAP improvement is 0.6998, with an average drop of 0.0894 in document classification performance. Additionally, concept gradient importance analyses and concept similarity heatmaps provide insights into the interpretability and structural separability of the GNN’s latent representations, achieved using CW.

Automatic Interpretation of Ancient Egyptian Texts for Education and Research

2025 · CHAPTER · en

We introduce a pipeline for interpreting Ancient Egyptian hieroglyphic texts combining OCR, transliteration, and translation. Designed for the low-resource data, our system improves accessibility for learners and efficiency for researchers. We evaluate its performance on a new diverse dataset reflective of real-world conditions.

MuMMy: Multimodal Dataset supporting VLM-based Egyptology Research Assistant

2025 · CHAPTER · en

We present the first multimodal dataset MuMMy, for developing research assistants that can interpret Egyptian hieroglyphic texts. It pairs images with Gardiner codes, transliteration, and English translation at two levels of granularity. We also evaluate several deep learning pipelines across OCR, transliteration, and translation tasks, revealing the complexity of the domain and the challenges posed by error accumulation.

Heterogeneous Graph Attention Networks for Scheduling in Cloud Manufacturing and Logistics

2024 · ARTICLE · en

Efficient task scheduling and resource allocation in manufacturing are vital for gaining competitive advantages in dynamic economic environments. Modern manufacturing systems must integrate logistics considerations such as delivery times and costs, yet traditional scheduling methods often overlook these factors. To address this gap, we investigate task scheduling in cloud manufacturing systems, emphasizing logistics integration. We propose a novel Graph Neural Network architecture for optimizing task scheduling by representing the problem on a heterogeneous graph, where nodes denote tasks and locations. Our model minimizes both manufacturing and logistics costs, achieving significant performance improvements over greedy algorithms and comparable results to strong genetic algorithms in large-scale scenarios with up to 20 locations. This work advances the efficiency and flexibility of cloud manufacturing systems, offering practical solutions for dynamic, cost-sensitive environments.

A Study of Graph Neural Networks for Link Prediction on Vulnerability to Membership Attacks

2024 · ARTICLE · en

Graph neural networks (GNNs) have shown great promise in a variety of tasks involving graph data, including recommendation systems. However, as GNNs become more widely adopted in practical applications, concerns have arisen about their vulnerability to adversarial attacks. These attacks can lead to biased recommendations, potentially causing economic losses and safety risks. In this work, we consider an industrial application of recommendation systems for transport lo- gistics and study their vulnerability to membership inference attacks. The dataset represents real train flows in Russia, published in the ETIS project. Experiments with three popular GNN archi- tectures show that all of them can be successfully attacked even when the adversary has minimal background knowledge. Specifically, an attacker with access to only 1-2% of the actual data can successfully train their own GNN model to infer the membership of a shipper-consignee associa- tion in the training set with an accuracy over 94%. Our study also confirms that overfitting is the primary factor that influences the attack performance of recommendation systems.

Device-Specific Facial Descriptors: Winning a Lottery with a SuperNet

2024 · CHAPTER · en