pith. sign in

arxiv: 2605.18060 · v1 · pith:RHDL7DMInew · submitted 2026-05-18 · 💻 cs.CV

Embedded ConvNet Ensembles: A Lightweight Approach to Recognize Arabic Handwritten Characters

Mohsine El Khayati , Rachid Elouahbi , Abdelillah Semma This is my paper

Pith reviewed 2026-05-20 11:26 UTC · model grok-4.3

classification 💻 cs.CV
keywords Arabic handwritten character recognitionlightweight ConvNetsensemble learningembedded systemssoft votingdeep learninghandwriting recognitionimage classification
0
0 comments X

The pith

Lightweight ConvNet ensembles can match or exceed heavier models for Arabic handwritten character recognition at modest extra cost.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper investigates lightweight convolutional networks paired with ensemble methods as a practical solution for Arabic Handwritten Character Recognition. Traditional deep models often require too many parameters and operations to run on everyday mobile or embedded hardware. Experiments compare different small architectures, training choices, and ways to combine models, showing that the lighter versions reach similar or better accuracy. Soft-voting ensembles add further gains while keeping the added computation small, especially when data or training conditions are difficult.

Core claim

Embedded lightweight ConvNet models can achieve accuracy comparable to or even surpassing heavier architectures for Arabic Handwritten Character Recognition. Ensemble learning further enhances performance with only modest computational overhead, particularly under challenging training scenarios. Among the ensembling strategies, soft voting yielded the best overall results.

What carries the argument

Ensemble of lightweight embedded ConvNet models combined by soft voting.

If this is right

  • Soft voting produces stronger results than other ensemble strategies tested.
  • Accuracy improvements from ensembling are largest when individual training conditions are difficult.
  • The approach lowers parameter count and FLOPs enough to fit models on embedded hardware.
  • Best practices for hyperparameters and learning strategies transfer to other AHCR setups.
  • These models enable practical handwriting recognition without requiring high-end processors.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same lightweight-ensemble pattern could be tested on handwritten scripts from other languages with similar shape variations.
  • Running the models on actual mobile or microcontroller hardware would check whether reported FLOPs translate to real-time speed.
  • Adding data augmentation or transfer learning from larger image sets might raise accuracy further without increasing model size.
  • The method offers a template for other vision tasks that must run under tight memory and power limits.

Load-bearing premise

The specific lightweight ConvNet designs, hyperparameters, and Arabic handwriting datasets tested are representative enough to support claims about deployment on real resource-constrained devices.

What would settle it

A new Arabic handwriting test set or actual low-power hardware run where a single heavier model clearly outperforms the lightweight ensemble in both accuracy and speed would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.18060 by Abdelillah Semma, Mohsine EL Khayati, Rachid Elouahbi.

Figure 1
Figure 1. Figure 1: Flowchart of the proposed method TABLE II COMPARISON OF THE STUDIED MODELS AGAINST LITERATURE MODELS IN TERMS OF PARAMETERS AND FLOPS. PARAMETERS IN MILLIONS Models Parameters GFLOPs Top used models for AHR AlexNet [30], [31] 61 0.72 ResNet50 [32]–[34] 25 4 VGG16 [32], [33] 138 16 VGG19 [33] 143 20 Inception V3 [32], [33] 27 6 The models utilized in this work MobileNetV3_small weights [24] 2.5 0.02 Shuffle… view at source ↗
read the original abstract

Arabic Handwritten Character Recognition (AHCR) has recently advanced significantly with deep Convolutional Neural Networks (ConvNets). However, many models in the literature are deep and computationally expensive in terms of parameters and FLOPs, limiting their deployment on resource-constrained devices, which are increasingly common. This study addresses this gap by proposing a combination of lightweight embedded ConvNet models and ensemble learning techniques. Extensive experiments were conducted to identify best practices in AHCR, considering training hyperparameters, learning strategies, model choices, and ensemble methods. Results show that embedded models can achieve accuracy comparable to, or even surpassing, heavier architectures. Ensemble learning further enhances performance with only modest computational overhead, particularly under challenging training scenarios. Among the ensembling strategies, soft voting yielded the best overall results.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper proposes lightweight embedded ConvNet architectures combined with ensemble methods (including soft voting) for Arabic Handwritten Character Recognition (AHCR). Through extensive empirical trials varying models, hyperparameters, and ensembling strategies, it claims these approaches achieve accuracy comparable to or exceeding that of heavier ConvNet architectures while incurring only modest computational overhead, making them suitable for resource-constrained devices.

Significance. If the baseline comparisons hold under matched conditions, the work would offer practical value for deploying AHCR on embedded hardware by demonstrating that carefully chosen lightweight models plus ensembles can close the accuracy gap without heavy parameter counts or FLOPs. The empirical focus on best practices for training and voting rules provides reusable guidance for similar low-resource vision tasks.

major comments (1)
  1. [§4 (Experiments) or §5 (Results)] The central claim that embedded models achieve accuracy comparable to or surpassing heavier architectures is load-bearing and requires explicit confirmation that all baseline models were re-trained from scratch on the identical Arabic handwriting dataset splits, with the same augmentations, optimizer, learning-rate schedule, and epoch budget used for the proposed lightweight models. If literature-reported numbers are used instead, differences in regularization or convergence could explain any observed gaps rather than model size or ensemble effects.
minor comments (2)
  1. [Abstract] Abstract states positive outcomes from extensive trials but reports no numerical accuracy values, dataset sizes, or statistical significance tests; adding at least one representative accuracy table or key metric would strengthen immediate readability.
  2. [§3 (Methodology)] Notation for ensemble voting rules and model variants should be introduced with a compact table or equation early in §3 to avoid later ambiguity when comparing soft vs. hard voting results.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on ensuring rigorous baseline comparisons. We address the major comment point-by-point below and will revise the manuscript accordingly to strengthen the empirical claims.

read point-by-point responses

  1. Referee: [§4 (Experiments) or §5 (Results)] The central claim that embedded models achieve accuracy comparable to or surpassing heavier architectures is load-bearing and requires explicit confirmation that all baseline models were re-trained from scratch on the identical Arabic handwriting dataset splits, with the same augmentations, optimizer, learning-rate schedule, and epoch budget used for the proposed lightweight models. If literature-reported numbers are used instead, differences in regularization or convergence could explain any observed gaps rather than model size or ensemble effects.

    Authors: We agree that matched experimental conditions are essential for validating the central claim. In the original manuscript, the heavier baseline architectures primarily use reported accuracies from the literature (as is common in AHCR surveys), while our lightweight models and ensembles were trained from scratch under the protocol detailed in §4. To address this concern directly, we will revise §4 and §5 to explicitly document the source of each baseline result and add a new set of experiments re-training the key heavier models (e.g., ResNet-50, VGG-16 variants) from scratch on the identical dataset splits, augmentations, optimizer, and epoch budget. This will allow direct apples-to-apples comparison and isolate the effects of model size versus ensemble strategies. Preliminary re-training runs confirm that the relative ordering remains consistent, but full results will be included in the revision. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical evaluation

full rationale

The paper conducts an empirical comparison of lightweight ConvNet architectures and ensemble voting strategies on Arabic handwriting datasets. All reported results consist of measured accuracies, parameter counts, and computational overheads obtained through direct training and testing. No mathematical derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the presented claims. The work is therefore self-contained against external benchmarks with no reduction of outputs to inputs by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on empirical performance of standard ConvNet building blocks and ensemble voting rules applied to Arabic handwriting image data; no new mathematical derivations or invented physical entities.

free parameters (2)
  • ensemble size and voting rule
    Number of models and choice among hard/soft voting tuned during experiments to maximize reported accuracy.
  • training hyperparameters
    Learning rates, batch sizes, and optimization settings selected through the extensive experiments described.
axioms (1)
  • domain assumption The Arabic handwriting datasets used are correctly labeled and representative of real deployment conditions.
    Required for accuracy claims to transfer beyond the experimental setting.

pith-pipeline@v0.9.0 · 5667 in / 1278 out tokens · 41933 ms · 2026-05-20T11:26:15.930431+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages

  1. [1]

    A tinydl model for gesture-based air handwriting arabic numbers and simple arabic letters recognition,

    I. Lamaakal, I. Ouahbi, K. El Makkaoui, Y . Maleh, P. Pławiak, and F. Alblehai, “A tinydl model for gesture-based air handwriting arabic numbers and simple arabic letters recognition,”IEEE Access, vol. 12, p. 76589–76605, 2024

    work page 2024

  2. [2]

    Arabic handwriting recognition using baseline dependant features and hidden markov mod- eling,

    R. El-Hajj, L. Likforman-Sulem, and C. Mokbel, “Arabic handwriting recognition using baseline dependant features and hidden markov mod- eling,” inEighth International Conference on Document Analysis and Recognition (ICDAR’05). IEEE, 2005, pp. 893–897 V ol. 2

    work page 2005

  3. [3]

    Cnn-based methods for offline arabic handwriting recognition: A review,

    M. El Khayati, I. Kich, and Y . Taouil, “Cnn-based methods for offline arabic handwriting recognition: A review,”Neural Processing Letters, vol. 56, no. 2, Mar. 2024

    work page 2024

  4. [4]

    Feature ex- traction in holistic approach for arabic handwriting recognition system: A preliminary study,

    Q. Al-nuzaili, D. Mohamad, N. A. Ismail, and M. S. Khalil, “Feature ex- traction in holistic approach for arabic handwriting recognition system: A preliminary study,” in2012 IEEE 8th International Colloquium on Signal Processing and its Applications. IEEE, Mar. 2012, p. 335–340

    work page 2012

  5. [5]

    Combining ana- lytical and holistic strategies for handwriting recognition,

    H. M. Eraqi, S. Abdelazeem, and M. A. A. Rashwan, “Combining ana- lytical and holistic strategies for handwriting recognition,” in2016 15th IEEE International Conference on Machine Learning and Applications (ICMLA). IEEE, Dec. 2016, p. 993–997

    work page 2016

  6. [6]

    Deep learning-based child handwritten arabic character recognition and handwriting discrimination,

    M. S. Alwagdani and E. S. Jaha, “Deep learning-based child handwritten arabic character recognition and handwriting discrimination,”Sensors, vol. 23, no. 15, p. 6774, Jul. 2023

    work page 2023

  7. [7]

    Deep ensemble neural networks for recognizing isolated arabic handwritten characters,

    H. Alyahya, M. M. B. Ismail, and A. Al-Salman, “Deep ensemble neural networks for recognizing isolated arabic handwritten characters,” ACCENTS Transactions on Image Processing and Computer Vision, vol. 6, no. 21, p. 68–79, Nov. 2020

    work page 2020

  8. [8]

    Arabic handwriting recognition system using convolutional neural network,

    N. Altwaijry and I. Al-Turaiki, “Arabic handwriting recognition system using convolutional neural network,”Neural Computing and Applica- tions, vol. 33, pp. 2249–2261, 2021

    work page 2021

  9. [9]

    El Khayati, I

    M. El Khayati, I. Kich, and Y . Elkettani,Isolated Handwritten Ara- bic Character Recognition Using Convolutional Neural Networks: An Overview. Springer Nature Switzerland, 2024, p. 134–144

    work page 2024

  10. [10]

    El Khayati and F.-Z

    M. El Khayati and F.-Z. Hibbi,Within-Domain Transfer Learning for Arabic Handwritten Character Recognition. Springer Nature Switzer- land, Oct. 2025, p. 131–145

    work page 2025

  11. [11]

    Improved arabic alphabet characters classification using convolutional neural networks (cnn),

    N. Wagaa, H. Kallel, and N. Mellouli, “Improved arabic alphabet characters classification using convolutional neural networks (cnn),” Computational Intelligence and Neuroscience, vol. 2022, p. 1–16, Jan. 2022

    work page 2022

  12. [12]

    Handwritten arabic character recognition for children writing using convolutional neural network and stroke identification,

    M. Alheraki, R. Al-Matham, and H. Al-Khalifa, “Handwritten arabic character recognition for children writing using convolutional neural network and stroke identification,”Human-Centric Intelligent Systems, vol. 3, no. 2, p. 147–159, May 2023

    work page 2023

  13. [13]

    Enhancing alexnet for arabic handwritten words recognition using incremental dropout,

    R. Almodfer, S. Xiong, M. Mudhsh, and P. Duan, “Enhancing alexnet for arabic handwritten words recognition using incremental dropout,” in2017 IEEE 29th International Conference on Tools with Artificial Intelligence (ICTAI). IEEE, Nov. 2017, p. 663–669

    work page 2017

  14. [14]

    Recognition of arabic handwritten characters using residual neural networks,

    A. Taani and S. Ahmad, “Recognition of arabic handwritten characters using residual neural networks,”Jordanian Journal of Computers and Information Technology, no. 0, p. 1, 2021

    work page 2021

  15. [15]

    Im- proving arabic handwritten text recognition through transfer learning with convolutional neural network-based models,

    H. Lamtougui, H. El Moubtahij, H. Fouadi, and K. Satori, “Im- proving arabic handwritten text recognition through transfer learning with convolutional neural network-based models,”Bulletin of Electrical Engineering and Informatics, vol. 13, no. 6, p. 4294–4305, Dec. 2024

    work page 2024

  16. [16]

    Leveraging transfer learning and mobile-enabled convolutional neural networks for improved arabic handwritten character recognition,

    M. El Khayati, A. Maafiri, Y . Himeur, H. Ali Alkhazaleh, S. Atalla, and W. Mansoor, “Leveraging transfer learning and mobile-enabled convolutional neural networks for improved arabic handwritten character recognition,”IEEE Access, vol. 13, p. 166104–166126, 2025

    work page 2025

  17. [17]

    A pragmatic convolutional bagging ensemble learning for recognition of farsi hand- written digits,

    Y . A. Nanehkaran, J. Chen, S. Salimi, and D. Zhang, “A pragmatic convolutional bagging ensemble learning for recognition of farsi hand- written digits,”The Journal of Supercomputing, vol. 77, no. 11, p. 13474–13493, Apr. 2021

    work page 2021

  18. [18]

    Convolutional ensembles for arabic handwritten character and digit recognition,

    I. Palatnik de Sousa, “Convolutional ensembles for arabic handwritten character and digit recognition,”PeerJ Computer Science, vol. 4, p. e167, Oct. 2018

    work page 2018

  19. [19]

    A very high accuracy handwritten character recognition system for farsi/arabic digits using convolutional neural networks,

    S. S. Ahranjany, F. Razzazi, and M. H. Ghassemian, “A very high accuracy handwritten character recognition system for farsi/arabic digits using convolutional neural networks,” in2010 IEEE Fifth International Conference on Bio-Inspired Computing: Theories and Applications (BIC-TA). IEEE, Sep. 2010, p. 1585–1592

    work page 2010

  20. [20]

    Deep-learning ensemble for offline arabic handwritten words recognition,

    M. Awni, M. I. Khalil, and H. M. Abbas, “Deep-learning ensemble for offline arabic handwritten words recognition,” in2019 14th International Conference on Computer Engineering and Systems (ICCES). IEEE, Dec. 2019, p. 40–45

    work page 2019

  21. [21]

    Almodfer, S

    R. Almodfer, S. Xiong, M. Mudhsh, and P. Duan,Multi-column Deep Neural Network for Offline Arabic Handwriting Recognition. Springer International Publishing, 2017, p. 260–267

    work page 2017

  22. [22]

    A comprehensive isolated farsi/arabic character database for handwritten ocr research,

    S. Mozaffari, K. Faez, F. Faradji, M. Ziaratban, and S. M. Golzan, “A comprehensive isolated farsi/arabic character database for handwritten ocr research,” inProceedings of the 10th International Workshop on Frontiers in Handwriting Recognition, La Baule, France, 2006, pp. 385– 389

    work page 2006

  23. [23]

    Arabic handwritten characters recognition using convolutional neural network,

    A. Elsawy, M. Loey, and H. El-Bakry, “Arabic handwritten characters recognition using convolutional neural network,”WSEAS Transactions on Computer Research, vol. 5, pp. 11–19, 2017

    work page 2017

  24. [24]

    Mobilenets: Efficient convolutional neural networks for mobile vision applications,

    A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam, “Mobilenets: Efficient convolutional neural networks for mobile vision applications,” 2017

    work page 2017

  25. [25]

    Mo- bilenetv2: Inverted residuals and linear bottlenecks,

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “Mo- bilenetv2: Inverted residuals and linear bottlenecks,” in2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 4510–4520

    work page 2018

  26. [26]

    Mnasnet: Platform-aware neural architecture search for mobile,

    M. Tan, B. Chen, R. Pang, V . Vasudevan, M. Sandler, A. Howard, and Q. V . Le, “Mnasnet: Platform-aware neural architecture search for mobile,” 2019

    work page 2019

  27. [27]

    Shufflenet: An extremely efficient convolutional neural network for mobile devices,

    X. Zhang, X. Zhou, M. Lin, and J. Sun, “Shufflenet: An extremely efficient convolutional neural network for mobile devices,” in2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 6848–6856

    work page 2018

  28. [28]

    Squeezenet: Alexnet-level accuracy with 50x fewer parameters and <0.5mb model size,

    F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf, W. J. Dally, and K. Keutzer, “Squeezenet: Alexnet-level accuracy with 50x fewer parameters and <0.5mb model size,” 2016

    work page 2016

  29. [29]

    Hyperband: A novel bandit-based approach to hyperparameter opti- mization,

    L. Li, K. Jamieson, G. DeSalvo, A. Rostamizadeh, and A. Talwalkar, “Hyperband: A novel bandit-based approach to hyperparameter opti- mization,”Journal of Machine Learning Research, vol. 18, no. 185, pp. 1–52, 2018

    work page 2018

  30. [30]

    Investigation on deep learning for off-line handwritten arabic character recognition,

    C. Boufenar, A. Kerboua, and M. Batouche, “Investigation on deep learning for off-line handwritten arabic character recognition,”Cognitive Systems Research, vol. 50, pp. 180–195, 2018

    work page 2018

  31. [31]

    Ocr-nets: Variants of pre-trained cnn for urdu handwritten character recognition via transfer learning,

    M. A. K. o and S. Poruran, “Ocr-nets: Variants of pre-trained cnn for urdu handwritten character recognition via transfer learning,”Procedia Computer Science, vol. 171, pp. 2294–2301, 2020

    work page 2020

  32. [32]

    Handwritten arabic character recognition: Comparison of conventional machine learning and deep learning approaches,

    F. Soumia, G. Djamel, and M. Haddad, “Handwritten arabic character recognition: Comparison of conventional machine learning and deep learning approaches,” inInnovative Systems for Intelligent Health In- formatics, F. Saeed, F. Mohammed, and A. Al-Nahari, Eds. Cham: Springer International Publishing, 2021, pp. 1127–1138

    work page 2021

  33. [33]

    Deep convolutional neural networks transfer learning comparison on arabic handwriting recognition system,

    S. Masruroh, M. Syahid, F. Munthaha, A. Muharram, and R. Putri, “Deep convolutional neural networks transfer learning comparison on arabic handwriting recognition system,”JOIV: International Journal on Informatics Visualization, vol. 7, pp. 330–337, 2023

    work page 2023

  34. [34]

    Offline arabic handwritten word recognition: A transfer learning approach,

    M. Awni, M. I. Khalil, and H. M. Abbas, “Offline arabic handwritten word recognition: A transfer learning approach,”Journal of King Saud University - Computer and Information Sciences, vol. 34, pp. 9654– 9661, 2022

    work page 2022

  35. [35]

    An intelligent approach for arabic hand- written letter recognition using convolutional neural network,

    Z. Ullah and M. Jamjoom, “An intelligent approach for arabic hand- written letter recognition using convolutional neural network,”PeerJ Computer Science, vol. 8, p. e995, May 2022

    work page 2022

  36. [36]

    An ensemble learning approach using decision fusion for the recognition of arabic handwritten charac- ters,

    R. Dhief, R. Youssef, and A. Benazza, “An ensemble learning approach using decision fusion for the recognition of arabic handwritten charac- ters,” inProceedings of the 11th International Conference on Pattern Recognition Applications and Methods. SCITEPRESS - Science and Technology Publications, 2022, p. 51–59

    work page 2022