Child handwriting enhancement poses unique challenges distinct from other document restoration tasks. Unlike the global degradation observed in historical manuscripts, children’s handwriting exhibits locally inconsistent faint strokes, misalignment with the writing line, and behavioral artifacts such as scribbles or partial erasures. When scanned, these irregularities are compounded by background noise, illumination variation, and scanner-induced distortions, deforming vital stroke information. Existing document restoration models, optimized for high-contrast ink or uniform degradations, tend to erase these subtle pencil traces, severely limiting their utility for downstream child handwriting analysis. To address these issues, we present TraceNet, an application-driven restoration network designed specifically for faint and irregular child handwriting. TraceNet integrates Swin Transformer-based self-attention within a single deterministic encoder-decoder restoration framework. Temporal diffusion embeddings are used as a conditioning signal to adapt restoration strength according to degradation severity, enabling selective recovery of faint strokes and line removal without over-smoothing. Evaluations on a private corpus of child handwriting show that TraceNet achieves strong performance across image-quality and downstream HTR metrics over state-of-the-art baselines. An additional evaluation on the DIBCO dataset further shows that the same framework is effective for historical document enhancement. Remarkably, the model maintains high restoration quality even under few-shot fine-tuning with only a small set of annotated samples, highlighting its efficiency and adaptability for real-world educational, archival, and behavioral handwriting analysis tasks.
Writer identification plays an important role in forensic document examination, authentication, and handwriting analysis. This study presents a hyperspectral imaging (HSI)-based writer identification framework that combines Principal Component Analysis (PCA), machine learning, and deep learning for robust writer recognition under previously unseen pen conditions. Experiments were conducted on two hyperspectral handwriting datasets: iVision HHID (54 writers; 512 × 650 × 149 cubes spanning 478–900 nm) and UWA WIHSI (7 writers; 480 × 752 × 33 cubes spanning 400–720 nm). Sauvola thresholding, binary image subtraction, and spectral segmentation were employed to isolate handwritten regions and extract ink-pixel spectral signatures. PCA was applied as a denoising and reconstruction technique, with hyperspectral cubes reconstructed using the first two principal components ( K=2 ) to enhance writer-discriminative spectral information while suppressing noise. The discriminative capability of the extracted features was further validated through ANOVA, where significant spectral bands exhibited high F-values and low p-values. Split-based experiments were first conducted to identify suitable classification models, with Linear Discriminant Analysis (LDA) and one-dimensional Convolutional Neural Networks (CNNs) demonstrating the strongest performance. These classifiers were subsequently evaluated using strict pen-level blind testing protocols. For the iVision HHID dataset, four train–test pen-ratio configurations (4:1, 3:2, 2:3, and 1:4) and an additional 1:5 protocol were considered, while the UWA WIHSI dataset was evaluated using the 4:1 pen-blind protocol. PCA-LDA achieved a maximum accuracy of 99.84
With the rapid development of computer vision technology, end-to-end text detection and recognition have effectively alleviated the error accumulation and suboptimal performance problems in traditional two-step methods. Due to the excellent modular characteristics of the two-step method, it is still widely favored in many competitions and practical scenarios. This paper proposes an innovative solution, a scene text detection and recognition algorithm based on connectors and multi-level adapters. It retains the core advantages of modularity while addressing error accumulation in the two-step method and improving performance. We use well-trained high-precision detectors and recognizers, locking them to preserve their learned capabilities. Then, a connector is used to link the detector and recognizer. Since the locked detector and recognizer cannot learn features through end-to-end optimization, we design multi-level adapters to facilitate feature learning between them. Experimental verification shows that the connector-based scene text detection and recognition method achieves significant results on multiple public datasets. The word-level end-to-end spotting score of this method reaches 69.3
Fine-grained sketch-based image retrieval (FG-SBIR) aims to retrieve the exact photo instance corresponding to a freehand sketch query. Despite recent progress, many existing methods rely primarily on global embedding alignment, while more recent approaches incorporate attention-based, transformer-based, and local matching strategies. However, these methods still struggle to handle the structural ambiguity and abstraction inherent in sketches. In this paper, we propose a structure-centric FG-SBIR framework that explicitly models cross-modal structural correspondence between sketches and photos. Instead of aligning holistic representations, both modalities are decomposed into structural primitives, and retrieval is formulated as an explicit structure-to-structure matching problem. To ensure geometric coherence, a structural consistency constraint is introduced to enforce topological alignment across modalities, while an uncertainty-aware weighting mechanism suppresses unreliable sketch structures caused by drawing variability. The proposed approach is evaluated on both category-level and fine-grained instance-level benchmarks, including Sketchy, TU-Berlin, QMUL Shoe-V2, and QMUL Chair-V2. Experimental results demonstrate consistent performance gains across multiple benchmarks, with particularly strong improvements on instance-level datasets, highlighting the effectiveness of explicit structural modeling for robust and discriminative sketch-based image retrieval.
The digitization of historical and folkloristic texts presents significant challenges for optical character recognition (OCR), particularly when documents contain complex layouts, embedded illustrations, irregular typography, or non-standard language. This study provides a systematic evaluation of six OCR approaches on two Slovene-language heritage collections: typewritten folklore manuscripts with uniform formatting, and visually heterogeneous issues of the children’s magazine Ciciban. The methods compared include Tesseract, Tesseract with GPT 5.2 post-processing, GPT 5.2 direct transcription, LLaMA 4 Maverick, Nanonets OCR-3, and Qwen-VL-OCR. Performance was assessed using character error rate, word error rate, and complementary sequence-based metrics against manually aligned ground truth. Results indicate that direct multimodal and document-oriented systems achieve the strongest accuracy on typewritten texts, while performance on Ciciban is more sensitive to layout structure. These findings highlight the document-sensitivity of OCR performance and point to the need for adaptive, content-aware pipelines that dynamically integrate multiple OCR strategies. To our knowledge, the study provides the first systematic benchmark of LLM-based OCR for Slovene folkloristic materials, offering practical insights for cultural heritage digitization.
Handwritten Text Recognition (HTR) remains a challenging task in document digitization, particularly for historical manuscripts written in low-resource languages such as Latin. In this paper, we focus on recognizing Latin texts from 16th–18th century manuscripts, which exhibit a wide range of handwriting styles. To address this, we propose AdapterTrOCR, a modular extension of the TrOCR model that incorporates two adapter modules: one for historical language adaptation and another for handwriting style adaptation. This architecture enables a robust transition from a modern English HTR model to one specialized in historical Latin. Given the limited availability of annotated data, we also explore Handwritten Text Generation (HTG) as a data augmentation strategy. Our results show the effectiveness of modular adaptation and synthetic data in improving HTR performance, achieving reductions in character error rate (CER) by 19.07 KU Leuven RDR .
Handwritten Text Recognition (HTR) is a challenging task in computer vision and the presence of crossed-out words in handwritten documents significantly impacts the recognition process. In this study, we first analyze the performance of three HTR methods on seven types of cross-outs. We compared HTR models based on Convolutional Recurrent Neural Networks (CRNN), Sequence-to-Sequence (Seq2Seq) architectures, and a Character Spotting (CSpot) approach. We then investigate two strategies to improve recognition of crossed-out words. The first explores improving the HTR models’ performance on different styles of cross-outs by adding crossed-out samples to the training data. The second explores improving HTR model performance by using a cross-out removal method to clean the handwriting. For our experiments, we use word images from the IAM dataset with synthesized cross-outs, as well as a newly created handwritten word image dataset with real cross-outs. The results show that both strategies significantly improve recognition accuracy; however, the best strategy depends on the context. CSpot performed better when trained with crossed-out samples. Cross-out cleaning works best for CRNN and Seq2Seq architectures, but the data requirement to implement this for real collections means it might not be practical for use. Datasets and code will be publicly available.
The rapid advancement of Multimodal Large Language Models (MLLMs) has opened new frontiers in automatic document understanding. However, understanding legal documents, with their unstructured nature and complex domain requirements, remains an underexplored challenge. We introduce IndianPCL, a large-scale document image dataset specifically curated for visual question answering (VQA) in the Indian legal context. IndianPCL comprises 358 multi page police complaint letters (250 printed and 108 handwritten) sourced from Indian police stations, over 6,300 annotated question–answer pairs for information localization and extraction. We use this dataset to systematically benchmark both proprietary and open-source MLLMs – including GPT-4o, Gemini 2.0 Flash, Phi-3.5 Vision-Instruct, LLaVA 1.6, LLaVA-Interleave, DeepSeek-VL 2 Tiny, Qwen-3-VL, and InternVl-3.5 – on two core tasks: standard VQA and Legal VQA, the latter requiring deeper legal reasoning and factual grounding. Results reveal substantial performance gaps: VQA accuracies range from 10-46
Historical document analysis has progressed to a point where the main bottleneck for many historical applications is not algorithms, but relevant interfaces, that can support historians’ workflow. While specialized tools exist for text processing and image search, we argue the community lacks a versatile collaborative platform enabling historians to analyze their own corpora from a particular perspective. As a step in this direction, we present aikon , a modular web-platform designed to empower historians with computer vision tools. aikon implements a complete workflow for historical document analysis, from corpus constitution to ai outputs validation and interpretation. It provides a comprehensive research environment combining source management tools with automated processing capabilities as well as multi-user validation and visualization interfaces. We showcase the potential of aikon by presenting modules enabling the investigation of graphical content transmission across large and diverse corpora, a problem for which few modern tools exist, and which is of key interest, for example in History of Science.
We present the first systematic benchmark evaluation of Optical Character Recognition (OCR) and Vision-Language Models (VLMs) for Turkish text recognition, addressing a critical gap in low-resource language processing. Turkish, with its agglutinative structure and unique characters (ç, ğ, ı, İ, ö, ş, ü), poses challenges for models trained on high-resource languages such as English. We developed a synthetic Turkish dataset of 13,200 images (6,600 from Atatürk’s Nutuk and 6,600 from Turkish Wikipedia) spanning three main text types: printed, handwritten, scene text. The dataset captures variations including Turkish character presence, word length effects, sentence vs. word recognition, and distortion types (rotation, resolution, noise, and blur). To assess generalizability, we additionally evaluate models on a Gold Standard set of 90 real-world images. We compare three model categories: traditional OCR systems, open-source VLMs, and commercial VLMs. Modern VLMs significantly outperform traditional OCR approaches, with GPT-4o and Qwen2.5 Vision-Language (Qwen2.5-VL) demonstrating superior performance. Images containing Turkish-specific characters posed significant challenges for all models, with only GPT-4o maintaining stable performance, highlighting the critical impact of training data composition on multilingual performance. Agglutinative word structure did not significantly affect recognition accuracy, whereas handwritten text recognition remains a persistent challenge across all systems. The open-source Qwen2.5-VL achieved comparable performance to commercial GPT-4o despite fewer parameters, showing strong potential as a computationally efficient alternative. This benchmark establishes a standardized evaluation framework for Turkish text recognition. We publicly release both datasets to enable reproducible research in low-resource language text recognition.
Handwritten text-line segmentation in historical manuscripts remains challenging due to degradation, overlapping strokes, and extreme data scarcity. We propose APAU-Net, a two-stage cascade architecture that improves Line Intersection-over-Union (Line IU) metric via learned anisotropic Gaussian priors. Stage 1 predicts topology-aware ellipsoidal priors from low resolution grayscale images using connected-component analysis and moment-based ellipse fitting. Stage 2 refines these priors at full resolution through a residual U-Net with adaptive per-pixel weighting and gated fusion. We evaluated APAU-Net on the challenging U-DIADS-TL (84 images, only 3 training pages per manuscript) and DIVA-HisDB benchmarks. It achieves an average Line Intersection-over-Union (Line IU) of 94.3
Traditional Optical Character Recognition (OCR) algorithms often struggle with documents that exhibit complex layouts, as they typically fail to model spatial relationships between textual elements. To address this limitation, we propose a novel architecture, termed LayoutGAT-OCR, which explicitly captures spatial dependencies between textual units using a Graph Attention Network (GAT) within a handwritten text recognition framework. In our approach, document elements (e.g., words or text regions) are represented as nodes in a graph, where each node is associated with visual features and positional encodings. These node representations are iteratively refined through graph attention mechanisms, enabling the model to incorporate contextual information from neighboring nodes and better understand layout-aware relationships. Experimental results demonstrate that LayoutGAT-OCR outperforms strong sequence-based baselines, including PARSeq, achieving a Character Error Rate (CER) of 9.85 on the challenging Qiaopi-HTR dataset, corresponding to a relative error reduction of over 49
Handwritten signatures are widely used bio-metrics for authenticating individuals using their unique traits. These systems commonly use deep convolutional Siamese networks to compare stroke-level patterns and decide the authenticity. Despite strong performance, these systems remain alarmingly vulnerable to adversarial perturbations: small, structured noise that can induce false acceptance or false rejections with minimal visible distortion. Although there are many adversarial attacks for images, audio, and text, most existing adversarial attacks cannot be applied to signature verification for three reasons. First, the foreground and background are clearly separated (requiring stroke-focused perturbations), Second, signatures are stroke-sparse and highly variable across writers and devices, and third, attackers typically lack perfect knowledge of the system in realistic black-box or white-box settings. Thus, generating effective, yet imperceptible, adversarial examples for signature verification remains unresolved. To address this gap, we present novel query-free universal black-box adversarial attack (QF-UBB) using Bayesian Dictionary Learning method. Prior attacks based on dictionary learning have shown promise, but they also suffer from limited adaptability, as traditional algorithms cannot generalize well to highly variable signature datasets where the problem is inherently dynamic. Our method has three key design principles. First, our algorithm generates a universal noise vector expressed as a sparse linear combination of stroke primitives learned by a Bayesian dictionary. Second, it is strictly restricted to signature foreground pixels (strokes). Third, the noise is produced without using any model gradients, classifier scores, or ground-truth labels making it query free completely black-box which does not require any information. Evaluations on three benchmark data sets show promising results in terms of attack success rate, imperceptibility, and transferability.
Training AI models to recognize characters engraved on medieval seals represents a challenging task in document analysis and cultural heritage preservation. Seal images often exhibit significant variations in style, material, state of conservation, and illumination, requiring large training sets that span a wide range of combinations. However, since collecting and annotating large sets of seal images is labor-intensive and time-consuming, synthetic training images are needed. In this work, we develop a pipeline to generate synthetic three-dimensional (i.e., computer-generated) seal images with random combinations of attributes. Our pipeline allows us to synthesize a large number of realistic, annotated, seal images under controlled combinations of attributes. We train different AI models for character classification, localization, and detection, enriching a small set of real seal or character images with a large number of synthetic ones either through fine-tuning or mixed training. Across all tasks and models, our experiments show that supplementing real samples with synthetic images boosts performance compared to training on real samples alone by margins topping 10
Offline Chinese handwritten text-line recognition (HCTR) remains challenging because standard CTC training on large-scale data is often dominated by abundant easy samples, while single-scale visual features are insufficient for simultaneously capturing fine-grained stroke details and higher-level structural semantics. To address these issues, we propose DA-MSFPN, an end-to-end difficulty-aware multi-scale recognition framework built on a CNN–Transformer backbone. DA-MSFPN constructs a multi-scale feature pyramid from hierarchical CNN representations and performs time-step-wise adaptive fusion via a lightweight gated attention module, while a two-stage sample-level difficulty-aware training strategy (SL-DAT) is introduced to improve hard-sample learning. On the CASIA-HWDB2.0–2.2 text-line benchmark, our method achieves 2.08
Document image restoration is a critical step in the preprocessing and analysis of digital images, aimed at recovering degraded document images. In recent years, generative models have achieved satisfactory results in the field of image restoration. However, the restoration results based on generative adversarial network methods exhibit issues such as color deviations, insufficient removal of overlapping regions, and unstable image quality. To address aforementioned issues, this paper proposes a document image restoration method based on diffusion models, specifically targeting the problem of restoring seen-through document images. This method follows a two-stage approach where the coarse stage focuses on restoring the primary contours of the foreground content, while the refinement stage handles detailed information. By introducing deep residual attention blocks and context-aware modules as two branches of the intermediate module, it is possible to better restore image details and enhance clarity. The proposed model combines pixel loss and diffusion loss to facilitate end-to-end training. Experimental results on the S-color0.5 dataset, DIBCO dataset, and MTDB dataset validate the effectiveness of the proposed method in improving color deviation issues and further enhancing the removal of overlapping regions seen-through.
Utilizing artificial intelligence for few-shot font generation (FFG) has become a trend in designing fonts for glyph-rich scripts. Most existing FFG approaches either globally disentangle the content and style of reference glyphs or decompose glyphs into strokes or radicals, then transfer these styles component-wise. However, they may fail to distinguish fine-grained local details or require predefined decomposition rules or special infeasible training strategies. This paper introduces a Hierarchical Quantization-based FFG approach (HQ-Font) by aggregating different-grained styles. It adopts a vector quantization strategy for glyph representation through unsupervised learning, enabling the contrasting and learning of discrete latent representations from low to high-level glyph feature spaces simultaneously without manual definition. A cross-attention mechanism is employed to transfer different granular styles of reference glyphs onto the discrete latent codes through contrastive learning, generating a complete set of content-agnostic style representations for different scripts. To this end, a font generation decoder performs hierarchical font synthesis, gradually mapping the corresponding global semantic and local stroke stylized codes from a given input to the final font image output. The number of glyph references as input can vary without the need for fine-tuning during testing, making HQ-Font more flexible. The experimental results demonstrate the effectiveness and generalization capability of HQ-Font across different scripts and also show its superiority when compared with other state-of-the-art FFG methods.
Unsupervised error estimation is examined under the framework of the statistical decision theory. The error probability of a classifier trained with a training set of a finite, fixed size is considered, along with an unsupervised naive estimate of this error. In general, the naive estimates tend to be significantly smaller than the empirical errors measured using ground truth class labels, but the estimated values can be easily calibrated with a function whose parameters can be trained using moderate amounts of class-labelled data. This way, true error rates can be accurately predicted for new unlabelled data. These ideas are applied to the essential classification problem which underlies the automatic transcription of text images. As a result, various methods are developed to predict the Word Error Rate (WER) of a text image recognizer, on unseen sets of images for which no ground truth transcripts are available. Experiments on three large handwritten text datasets show that the error rates predicted by some of these methods are sufficiently accurate for practical applications. More specifically, absolute deviations of the predicted word error percentages from the corresponding real WER are lower than 1.7
Answer sheet layout analysis is an important preliminary task for handwritten answer sheet recognition. Existing layout analysis methods can analyze the handwritten content of students in the handwritten answer sheet to the level of text lines. However, because of the presence of various elements such as text, mathematical expressions, and systems of equations within the answer sheet text lines, their content is heterogeneous and structurally complex, making direct recognition impossible. To address this challenge, we propose an efficient multi-scale bidirectional parallel feature fusion network (EMBFN). This model enables the analysis of handwritten answer sheet text lines and optimizes feature fusion by enhancing the representations extracted by the backbone network. It mitigates feature conflicts arising from interlayer discrepancies during fusion, thereby reducing errors in sparse boundary analysis and the misclassification of similar elements. EMBFN uses a dynamic competitive fusion block to address detail ambiguity and semantic conflicts arising from static fusion strategies, thereby enhancing the effectiveness of feature fusion. In addition, we constructed an Answer Sheet Text Line Analysis (ASTA) dataset to complement missing analysis elements, such as logical symbols and smudges, in existing public datasets. It contains handwritten lines of text from a variety of disciplines and annotates all elements within the lines of text. Experimental results on both the ASTA dataset and publicly available datasets demonstrate the effectiveness of our proposed method. Notably, our network model contains only 2.2 million parameters, while achieving an accuracy of 96.7
The Ethiopic script used in the Amharic Language presents persistent challenges for Optical Character Recognition (OCR) due to its large character vocabulary, diacritics, and high variability in handwriting. Although Amharic is spoken by over 58 million people, progress in OCR has been constrained by the lack of large, diverse, sentence-level datasets. Existing datasets are small, synthetic-only, or limited to character- or word-level annotations, preventing models from capturing the complexity of real documents. We introduce Fidel, the first large-scale Amharic OCR dataset spanning handwritten, typed, and synthetic text. Fidel contains 40k handwritten and 28k typed line images collected from 411 native writers, providing broad coverage of handwriting styles and modern vocabulary. We further formalise our approach as a scalable data acquisition and preprocessing pipeline, deskewing, line extraction, and alignment, designed to guide future dataset creation for low-resource scripts. To complement the real data, we generate high-quality synthetic Amharic text images to support robust model training. Using Fidel, we construct the first comprehensive benchmark for Amharic OCR, evaluating seven deep learning based OCR models. These models span CNN, CTC, transformer and hybrid architectures, enabling a robust assessment of domain transfer and modality-specific performance across handwritten, typed, and synthetic text. The best-performing model trained on Fidel achieves state-of-the-art results, with a CER of 2.64
