2026· International Journal of Applied Science and Research· Vol 09, pp. 01-19· 0 citations
TL;DR
The paper will suggest and mathematically model a confidence-aware hybrid retrieval system in Arabic document images with the use of keyword queries, and presents Arabic-oriented framework, reproducible evaluation protocol, measured auxiliary evidence and the identification of the dedicated Arabic retrieval benchmark that will be needed to complete target-domain validation.
Abstract
The high rate of digitization of Arabic institutional archives has created masses of page images, which are hard to search where good textual metadata are unavailable. Lexical indexing can be supported by Optical Character Recognition (OCR), but its reliability suf ers from blur, skew, low contrast, compression artifacts, complex layout, and Arabic-specific script features like contextual letter forms, ligatures and dots, and optional diacritics. The paper will suggest and mathematically model a confidence-aware hybrid retrieval system in Arabic document images with the use of keyword queries. The framework integrates a lexical branch (OCR-based), a visual region-matching branch, branch-based score calibration, confidence gated fusion, re-ranking and keyword localization. The empirical part is limited by the measures which are directly justified by the measured datasets. On the FUNSD testing split (50 pages; 668 normalized single-token searches chosen in ground-truth annotations) an experiment of keyword retrieval at the page level was implemented. Using Tesseract 5.5.0 English OCR and BM25 indexing, OCR-BM25 achieved P@1 = 0.867, mAP = 0.678, NDCG@10 = 0.713, and MRR = 0.888; the confidence-weighted OCR variant achieved P@1 = 0.870, mAP = 0.672, NDCG@10 = 0.708, and MRR = 0.890. The given measured values confirm the auxiliary lexical retrieval protocol of non-Arabic noisy forms and demonstrate that raw OCR confidence is not uniformly beneficial when it comes to ranking metrics. They are not presented as Arabic hybrid retrieval performance. The assessed IFN/ENIT files include Arabic handwritten word images and segmentation XML but do not include page-level keyword relevance labels, whereas assessed RVL-CDIP test folders include document-class images instead of keyword annotations. It is based on this that the paper presents Arabic-oriented framework, reproducible evaluation protocol, measured auxiliary evidence and the identification of the dedicated Arabic retrieval benchmark that will be needed to complete target-domain validation.
English, Hindi and Marathi Multilingual optical character recognition (OCR) is challenging where the text is scanned on real-life images due to blur, uneven lighting, background noise, and mixed script writing. The paper introduces a lightweight script-aware OCR system that enhances recognition accuracy by using coordinated preprocessing, connected-component-based word localization, script routing, and confidence-based re-recognition instead of retraining the entire system. The pipeline consists of bilateral filtering, fast non-local means denoising, contrast-limited adaptive histogram equalization (CLAHE), adaptive thresholding, connected component analysis, script identification, and dual-pass Tesseract recognition. The algorithm is aimed at low-resource multilingual document comprehension in which interpretability, tunability, and relatively low computational efficiency are of significance. Experiments using a 5000 image corpus indicate a recognition accuracy of 93.0 on average and language-wise accuracies of 95% in English, 92% in Hindi and 90% in Marathi. The error rate in the characters within English, Hindi and Marathi were 4.2, 6.5 and 7.1 respectively and the script identification stage was 0.94 with a recall 0.93 and F1-score 0.935. The analysis of ablation reveals that adaptive thresholding and script identification are the most significant ones. The findings suggest that a well-designed, script-sensitive pipeline offers both a viable and technically rational solution to multilingual OCR of realistic images.
P. Patil, Monali Y.Khachane· 2026 International Conferenc...· 0 citations
In computer vision, word detection in images is still a major challenge, especially for applications like scene interpretation, document indexing, and machine translation. Arabic script presents particular difficulties because of its cursive style, the use of diacritical marks, and the range of typefaces and orientations, which sometimes lead to errors in standard models, even though this effort has achieved great strides in Latin scripts. In this work, we propose a two-step method for Arabic Text Detection (ATD), aiming to improve accuracy without overly complex pipelines. Our approach starts with a set of heuristic rules to pre-select candidate regions. These rules rely on basic geometric and statistical cues, which helped us eliminate much of the irrelevant background noise in initial tests. To further improve the choices, we use an EfficientNet-B1 CNN. This model uses Global Average Pooling and a final Softmax layer to distinguish between text and non-text regions. The evaluation, which was conducted on a broad dataset of Arabic scene photos, revealed that combining heuristics with CNN verification enhances results significantly. For instance, we observed an increase in detection accuracy from 89.45% (heuristics alone) to 96.46% with the full pipeline. Moreover, the CNN classifier itself reached an accuracy of 97.77%. These findings confirm the robustness of our approach, particularly in cluttered visual environments.
A. Amrouche· Iraqi Journal of Science· 0 citations
Experimental results show that the proposed model can not only recognize characters in document images but also locate word boundaries, removing the need for an extra word segmentation step in a conventional sequential pipeline.
Marry Kong, Rina Buoy, Sovisal Chenda et al.· 0 citations
Scene text detection and recognition in multilingual environments remains challenging, particularly for morphologically complex scripts such as Arabic. This paper presents an end-to-end deep learning system for detecting and recognizing Arabic and English text in natural scene images. The detection pipeline combines a Swin Transformer Tiny backbone pretrained on ImageNet-22K, a Feature Pyramid Network for multi-scale fusion, and a DBNet++ head, achieving an F1-score of 94.52% on a strictly held-out 2,000-image ICDAR 2019 MLT test set. For recognition, PARSeq with permutation language modeling is trained on 656,868 cropped word samples (85/15 split), reaching 89.50% word accuracy and 94.13% character accuracy on an 865-character bilingual charset. The system includes a vertical-projection word-segmentation fallback and a smart RTL/LTR ordering algorithm. Ablations show FPN contributes +4.6 pp and DBNet++ +3.22 pp F1. Comparisons with prior baselines are reported as non-comparable references. The contribution is a fully reproducible bilingual Arabic-English pipeline.
Taher Ali Mahmood, Y. Mohammad· IEEE Jordan Conference on Ap...· 0 citations
Offline handwritten word recognition has progressed from whole-word classification to sequence transcription, yet many systems depend on large annotated corpora and exploit lexical regularities over explicit character evidence. This paper presents an alternative formulation for Arabic offline handwritten word recognition, treating characters as spatial objects detected via a Faster Region-Based Convolutional Neural Network rather than symbols generated by a one-dimensional decoder. We construct and release a character-level annotated subset of 2153 handwritten word images from a standard Arabic benchmark, exporting matched detection, sequence, and word-class labels. We also introduce an open-source subword exchange toolkit that creates a controlled structural-generalization benchmark by swapping subwords while preserving handwriting style. Experiments compare the proposed detector against whole-word and sequence-based baselines on both the original held-out split and the perturbed benchmark. Results show sequence models degrade sharply under structural recombination, whereas the proposed detector remains stable, achieving a 26.56% character error rate and 70.0% word accuracy on the perturbed benchmark. These findings demonstrate that explicit character localization provides a robust, data-efficient alternative for Arabic handwritten text recognition in low-resource settings.
Sofiane Medjram, Ruwaidah Saud Alnejaidi· Applied Sciences· 0 citations
Existing approaches to building line-level Arabic handwritten-text-recognition (HTR) training data either rely on fully manual annotation, which does not scale, or on automatic OCR-to-reference alignment methods not yet extended to multi-script, two-zone (main-plus-margin) manuscript layouts with a provable correctness guarantee. We present RefLAM (Reference-grounded Line Annotation for Manuscripts), a pipeline converting manuscript page images and clean transcriptions into validated, line-level ground truth without sacrificing human oversight. RefLAM couples a deep-learning page-segmentation model with a multimodal large language model (MLLM) for structured OCR and a diacritic-agnostic fuzzy alignment engine that grounds each OCR line in a contiguous span of the reference text, with a character-level confidence score in $[0,100]$. A perfect score is provably equivalent to character-for-character identity of the normalised strings (the Confidence-100 rule), verified with no counterexample across the released corpus. A reviewer can thus trust a perfect score, confirming most lines at a glance rather than retyping them, so annotation becomes triaged, with attention concentrated on uncertain alignments. Across 7 fully page-validated books we measured a 75$\times$ throughput gain over manual annotation (3,000 vs. 40 lines/hr); applying the same guarantee to 7 further books, we retained 16,533 confidence-100 main-text lines within one week, excluding sub-100 lines rather than manually correcting them. Using RefLAM, we release AraMS-28k: 14 historical Arabic manuscript books, 3,043 pages, and 27,971 main-text and 629 margin-line annotations with bounding boxes, layout labels, and insertion anchors for 191 margin entries (30.4%). We also finetune Muharaf-pretrained baselines (including HATFormer) on AraMS-28k and report CER results confirming its practical utility for downstream HTR training.
