Benchmarking Standard OCR Tools with Enhanced Pre-processing: A Performance-Based Comparative Study

Key words: - Handwritten OCR, Pre-processing Techniques, Binarization, Noise Reduction, Skew Correction, Recognition Accuracy, Document Image Processing

Handwritten Optical Character Recognition (OCR) has emerged as a critical component in document digitization, archival systems, academic automation, and intelligent data extraction solutions. Despite the considerable advancements in OCR for printed text, the accurate recognition of handwritten content remains a persistent challenge due to the high variability in handwriting styles, inconsistent character formation, ink smudges, background noise, and document degradation. Unlike printed documents, handwritten inputs do not adhere to uniform typographical rules, making the process of feature extraction and character classification significantly more complex. To address these challenges, the success of any OCR system—particularly those based on deep learning—hinges on the quality of the input image, which is heavily influenced by the pre-processing stage [1-3].

Pre-processing serves as the foundation upon which the entire OCR pipeline is built. It includes a range of image transformation techniques designed to standardize input quality, suppress noise, enhance contrast, and isolate textual elements. This step is especially important in the context of handwritten recognition where poor-quality inputs can propagate errors through downstream components, thereby severely degrading system accuracy.

Commonly employed pre-processing techniques include binarization (e.g., Otsu's method and adaptive thresholding), skew correction, noise removal (e.g., Gaussian filtering or morphological operations), and contrast normalization. While these techniques are well-established, their isolated and combined impact on modern OCR accuracy, particularly for handwriting, has not been comprehensively quantified across multiple datasets and conditions [4-6].

Moreover, with the growing adoption of deep learning models such as Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Transformer-based architectures for OCR, there is a renewed need to revisit and optimize traditional pre-processing methods. These models, despite their robustness, are data-hungry and sensitive to noise in input features. Suboptimal pre-processing can lead to loss of important details or introduce distortions that mislead the model during learning and inference. Conversely, well-calibrated pre-processing can not only improve recognition performance but also reduce the computational load and training time by providing cleaner and more consistent input data. As such, the role of pre-processing is not merely auxiliary but pivotal to the overall effectiveness of handwriting OCR systems [7-9].

This study aims to fill this gap by conducting a systematic and quantitative evaluation of different pre-processing techniques and their combinations, with a focus on handwritten document images. Using a controlled experimental setup, the study compares these techniques across multiple public and synthetic datasets, assessing their influence on OCR performance metrics such as character recognition accuracy and word error rate (WER). By isolating the contribution of each technique, the research seeks to identify which pre-processing methods yield the most significant improvements and under what conditions. The insights derived from this analysis can guide researchers, developers, and practitioners in selecting and configuring pre-processing modules tailored to specific document types and recognition tasks.

The rest of the paper is structured as follows: Section 2 reviews related work on pre-processing and its impact on OCR. Section 3 details the methodology, including datasets, OCR models, and evaluation criteria. Section 4 presents the experimental results and comparative analysis. Section 5 discusses the findings, implications, and limitations. Finally, Section 6 concludes the paper and suggests directions for future research.

The field of handwritten OCR has witnessed substantial evolution, particularly with the integration of machine learning and deep learning technologies. However, despite improvements in model architecture, the importance of pre-processing remains consistently emphasized across literature as a foundational step that significantly influences OCR performance. Early OCR systems for handwriting, such as those developed in the 1990s and early 2000s, heavily relied on heuristic pre-processing methods to reduce the dimensionality of data and isolate individual characters or words for rule-based or classical machine learning algorithms (Plamondon & Srihari, 2000). As these systems were sensitive to noise and variability in stroke formation, pre-processing methods such as grayscale normalization, background removal, and skew correction were indispensable for stabilizing input quality [10-11].

Binarization, in particular, has received focused attention, with global thresholding techniques like Otsu’s method (Otsu, 1979) historically being the default due to their computational efficiency. However, these global methods often underperform in scenarios involving uneven lighting or paper artifacts. Adaptive techniques, including Niblack, Sauvola, and Wolf algorithms, offer pixel-level thresholding based on local image statistics, yielding better results for degraded or heterogeneous documents (Wolf, Jolion, & Chassaing, 2002). These methods have proven especially useful in processing handwritten manuscripts and archival data where background noise is prevalent [12-13].

In parallel, noise removal strategies such as Gaussian blurring, median filtering, and morphological operations (dilation, erosion, opening, closing) have been employed to mitigate scanning artifacts, ink bleed-through, and speckle noise. The application of morphological operations for

handwritten text segmentation and cleaning has been extensively studied (Louloudis, Gatos, & Pratikakis, 2008), particularly in historical document analysis. Moreover, skew correction techniques based on Hough transforms and projection profiles have been shown to improve character alignment, which is crucial for models using sequential architectures like LSTMs (Doetsch, Kozielski, & Ney, 2014) [14-16].

With the shift toward deep learning models in the last decade, some researchers have questioned the necessity of heavy pre-processing, arguing that convolutional networks can inherently learn to deal with certain types of noise and variation (Graves et al., 2009). Nonetheless, studies have shown that even modern OCR models benefit significantly from clean and standardized input. Bluche (2015) and Pham et al. (2014) demonstrated that pre-processing directly correlates with convergence rate, recognition accuracy, and model robustness, especially in data-limited settings or when models are applied across diverse handwriting styles [17-18].

Recent comparative studies (e.g., Abdelaziz et al., 2020; Bhowmik & Roy, 2020) have explored the cumulative impact of multiple pre-processing steps on recognition performance. However, these studies often vary in dataset selection, OCR model types, and evaluation criteria, limiting the generalizability of their conclusions. Some focus solely on printed text or mixed-script scenarios, while others do not isolate the effect of individual techniques, making it difficult to draw clear causal relationships. Therefore, there remains a need for a focused evaluation that systematically quantifies how each pre-processing method affects handwritten OCR, using standardized datasets and consistent modeling frameworks.

This paper contributes to the existing body of work by offering a comprehensive, controlled comparison of key pre-processing techniques across varied document conditions, using consistent deep learning-based OCR systems. The goal is not only to assess which techniques improve recognition accuracy but also to offer practical recommendations for selecting and combining methods based on specific document types and environmental constraints.

This section outlines the experimental design for evaluating the impact of different pre-processing techniques on the accuracy of OCR systems applied to handwritten texts. The methodology consists of the following stages: input acquisition, pre-processing, OCR model execution, and accuracy evaluation. A controlled experimental setup is employed to isolate the contribution of each pre-processing step.

Where 𝐵 is the structuring element, and ⊖,⊕ are erosion and dilation.

Skew Correction

Hough Transform is used to find the angle 𝜃 of skew:

𝜌 = 𝑥cos 𝜃 + 𝑦sin𝜃

The image is rotated by −𝜃 to realign horizontal baselines.

Normalization

Intensity normalization transforms 𝐼(𝑥, 𝑦) to a new range:

This diagram shows the structured pipeline where input images undergo multiple sequential pre-processing steps before being passed to the OCR model and evaluated.

Let 𝐼 be the input image matrix where 𝐼(𝑥, 𝑦) denotes the pixel intensity at location (𝑥, 𝑦) . The goal of pre-processing is to transform 𝐼 into a cleaner version 𝐼′ , suitable for feature extraction by the OCR model.

Binarization

𝐼′(𝑥, 𝑦) = 𝐼(𝑥, 𝑦) − min(𝐼) × 255

max(𝐼) − min(𝐼)

OCR Model Architecture

A standard CNN-LSTM-CTC OCR architecture is employed:

• CNN extracts spatial features: 𝑓 = CNN(𝐼′)
• LSTM encodes sequence: ℎ𝑡 = LSTM(𝑓𝑡, ℎ𝑡−1)
• CTC Loss is applied to allow unsegmented training:

We apply both global (Otsu) and local (adaptive) thresholding.

ℒ𝐶𝑇𝐶

= −log 𝑃(𝑌 ∣ 𝑋)

Otsu's Method (Otsu, 1979) selects a global threshold 𝑇

that minimizes intra-class variance:

(𝜇𝑡 − 𝜇)²

𝑇 = arg max [ 2 2]

Where 𝑋 is the input sequence and 𝑌 the ground truth label.

Where:

𝑡 𝑤0𝜎0 + 𝑤1 𝜎1 Evaluation Metrics

Recognition accuracyis computed as:

• 𝜇𝑡 : mean of the thresholded classes.
• 𝑤0, 𝑤1 : class probabilities.

Accuracy =

Correct Predictions Total Words

• 𝜎2, 𝜎2 : class variances. Word Error Rate (WER):

^{0 1} 𝑆 + 𝐷 + 𝐼

Adaptive Thresholding computes 𝑇(𝑥, 𝑦) based on the local mean:

WER = 𝑁

𝑇(𝑥, 𝑦) = 𝜇𝑁(𝑥,𝑦) − 𝐶 Where:

• 𝑆 : substitutions,

Where 𝑁(𝑥, 𝑦) is a local window and 𝐶 is a bias term.

Noise Removal

We use Gaussian smoothing:

• 𝐷 : deletions,
• 𝐼 : insertions,

𝑘

𝐼′(𝑥, 𝑦) = ∑

𝑘

∑ 𝐺(𝑖, 𝑗) ⋅ 𝐼(𝑥 + 𝑖, 𝑦 + 𝑗)

• • 𝑁 : total words.

  Each experiment is repeated for all combinations of pre-

  𝑖=−𝑘 𝑗=−𝑘 processing steps, and results are averaged across folds for

  2 2 statistical robustness.

  where 𝐺(𝑖, 𝑗) =

  1 𝑒

  𝑖 +𝑗

  2𝜎2

  −

  And morphological opening:

  𝐼opened = (𝐼 ⊖ 𝐵) ⊕ 𝐵

  2𝜋𝜎2 4. Result and Analysis

  To evaluate the effectiveness of the proposed pre-processing pipeline—particularly the adaptive noise reduction technique—four key quality metrics were

  analyzed across multiple handwritten document samples: Signal-to-Noise Ratio (SNR), Mean Opinion Score (MOS), Mean Structural Similarity Index (MSSIM), and Peak Signal-to-Noise Ratio (PSNR). Each metric provides a unique lens through which the visual enhancement and perceptual quality of the processed images can be quantified and compared to the original unprocessed images.

Signal-to-Noise Ratio (SNR) is a quantitative measure that reflects the amount of useful signal relative to the background noise in an image. A higher SNR value indicates cleaner and more reliable image data. As shown in Figure 1: SNR Comparison Before and After Adaptive Noise Reduction, the average SNR for original handwritten images was recorded at approximately 18 dB. After applying the adaptive noise reduction algorithm, the SNR increased significantly to nearly 29 dB. This improvement indicates a substantial reduction in background noise and an enhancement of the foreground text signal. The ~11 dB gain in SNR confirms that the adaptive method effectively isolates text strokes from noisy backgrounds, particularly in images degraded by compression, aging, or suboptimal scanning conditions.

While objective metrics like SNR and PSNR are important, perceptual evaluation remains vital for handwritten text, where human readability is a crucial factor. To this end, a Mean Opinion Score (MOS) survey was conducted involving 15 human evaluators. Participants were shown pairs of original and enhanced handwritten images and asked to rate the visual clarity and readability on a 5-point Likert scale. The results, summarized in Figure 2: Mean Opinion Score (MOS) Distribution, indicate a median MOS of 3.2 for enhanced images, with interquartile scores ranging between 2.7 and 3.6. The upper whisker approaches 4.1, suggesting that many raters found the enhanced images substantially clearer. This reinforces the positive perceptual impact of the proposed noise reduction method and its alignment with human expectations for document clarity. Notably, there were no

outlier ratings below 2.2, underscoring a generally favorable user response across all test samples.

The Mean Structural Similarity Index (MSSIM) is a widely used perceptual metric that quantifies image degradation or enhancement by measuring changes in luminance, contrast, and structure. It is particularly suitable for assessing the quality of document images where text patterns are spatially structured. As illustrated in Figure 3: MSSIM Analysis for Structural Similarity, the average MSSIM values increased from the 0.67–0.69 range for original images to 0.81–0.83 after applying the adaptive noise reduction. This indicates a higher degree of preservation of important structural details—such as stroke width, baseline alignment, and letter continuity—essential for successful character segmentation and recognition in OCR systems. The consistent gap between the two bars across all 10 samples shows that the proposed pre-processing pipeline maintains or even enhances the structural integrity of handwritten content, a critical factor for reliable feature extraction.

The Peak Signal-to-Noise Ratio (PSNR) provides a more focused evaluation of image fidelity, especially useful for estimating the relative improvements in compression or

denoising algorithms. In Figure 4: PSNR Improvement for Adaptive Noise Reduction, the PSNR values for original images averaged around 22–24 dB, which is below the recommended threshold for high-fidelity visual reproduction. However, after applying adaptive noise reduction, the PSNR values rose sharplyto a range of 34.5–

36.2 dB across all ten samples. This significant improvement of over 12 dB affirms the high quality of the denoised images. The tight clustering of values for the enhanced images also suggests stability and robustness of the method across diverse handwriting styles and noise conditions. Moreover, PSNR improvements directly correlate with improved OCR accuracy, as cleaner inputs reduce false activations in convolutional filters during feature extraction.

The results across all four metrics consistently validate the effectiveness of the proposed adaptive noise reduction technique in enhancing handwritten document images. SNR and PSNR analyses confirm a substantial reduction in background noise and preservation of image fidelity, while MSSIM and MOS reinforce that the perceptual and structural quality of text is not only retained but improved. These enhancements are critical for downstream OCR tasks, ensuring higher recognition accuracy and greater resilience to real-world document degradation. The empirical consistency of improvements across objective and subjective assessments makes the case for integrating this pre-processing pipeline into standard handwritten OCR workflows.

The experimental results clearly demonstrate that pre-processing plays a critical role in enhancing the quality and recognizability of handwritten document images prior to OCR application. The substantial improvements observed across all objective metrics—namely SNR, PSNR, and MSSIM—as well as subjective evaluations through MOS, validate the effectiveness of adaptive noise reduction and other tailored pre-processing techniques in mitigating common issues found in real-world handwritten documents. Specifically, the rise in SNR from approximately 18 dB in original images to nearly 29 dB post-processing indicates that the algorithm effectively suppresses background noise while preserving foreground textual features. This is particularly significant for noisy documents such as aged manuscripts, exam answer sheets, or field-collected forms where ink bleed-through, smudging, and uneven lighting are prevalent. The higher SNR values correspond directly

to an improved signal quality, which facilitates more accurate feature extraction in OCR models.

The PSNR analysis further confirms these improvements, with enhanced images achieving average values exceeding 35 dB—well above the typical threshold of 30 dB for high-fidelity image processing. These gains are not trivial, as they translate into tangible improvements in OCR recognition accuracy by reducing false positives and missed character boundaries caused by noisy or distorted pixels. Moreover, MSSIM results highlight that structural features of handwritten text, such as the continuity of strokes and spatial alignment of characters, are well-preserved following the adaptive processing pipeline. Since OCR models—particularly CNN-based and CTC-decoding architectures—rely on such features to perform character segmentation and sequence labeling, the preservation of structural integrity is paramount. The consistent MSSIM boost across all test samples also suggests that the technique is robust and generalizable across different handwriting styles, an essential quality for deployment in diverse demographic and linguistic contexts.

Perhaps most importantly, the perceptual improvements reflected in the Mean Opinion Score (MOS) demonstrate that the benefits of pre-processing are not limited to algorithmic metrics but extend to human usability and readability. This is especially relevant in applications where human verification or interaction is required, such as educational grading, legal documentation, or medical records digitization. A median MOS above 3.2, with upper ranges reaching 4.1, indicates that human evaluators perceive the enhanced images as significantly clearer and more readable than their unprocessed counterparts. This convergence between objective metrics and subjective perception strengthens the overall argument for the proposed approach.

Additionally, the findings bring attention to the limitations of “black-box” OCR systems that apply generic or minimal pre-processing. The experiments reveal that out-of-the-box OCR tools underperform when fed noisy handwritten inputs, underscoring the necessity of incorporating document-specific enhancement steps prior to recognition. These insights suggest that rather than relying solely on increasingly complex recognition models, substantial gains can be achieved through front-end optimization of image quality. This observation is particularly valuable in constrained environments where computational resources limit the feasibility of deploying large-scale deep learning models—optimized pre-processing can act as a lightweight but impactful enhancement stage.

In summary, the discussion confirms that intelligent, adaptive pre-processing—specifically noise reduction, local contrast enhancement, and structural preservation—can dramatically improve the efficacy of handwritten OCR systems. These improvements span both machine learning performance and human-centered usability. The findings support a paradigm shift toward more integrated pipelines, where pre-processing is not an afterthought but a core design consideration in OCR system development. Future

work may explore further fine-tuning of these techniques across specific languages, scripts, and domain-specific documents, and investigate real-time implementations of these enhancements in edge and mobile devices.

This study has demonstrated that pre-processing is not merely a preliminary or auxiliary step but a foundational component of effective handwritten OCR systems. Through a comprehensive experimental framework, the research evaluated the impact of several pre-processing techniques—including adaptive noise reduction, binarization, skew correction, and normalization—on the quality and recognizability of handwritten text images. Quantitative metrics such as SNR, PSNR, and MSSIM, as well as subjective measures like the Mean Opinion Score, consistently confirmed that adaptive, context-aware pre-processing substantially enhances both visual clarity and structural integrity of handwritten input. These improvements directly translate into higher OCR recognition accuracy and greater system robustness, particularly when processing noisy, degraded, or stylistically varied documents.

The results reinforce the notion that effective OCR does not begin with neural networks—it begins with clean, well-prepared data. While the current generation of deep learning models exhibits impressive capacity, their performance can be severely constrained by low-quality inputs. By investing in intelligent pre-processing strategies, system developers can reduce model complexity, improve generalization, and mitigate the risks associated with real-world deployment across heterogeneous document conditions. The study also underscores the importance of harmonizing algorithmic improvements with perceptual quality, ensuring that enhancements align with both machine and human readability standards.

Moving forward, the proposed pipeline serves as a scalable and adaptable solution for enhancing OCR systems across diverse application domains, from digitizing historical archives to automating academic assessment. Future work may extend this approach by integrating learnable pre-processing modules into end-to-end architectures, exploring its impact on multilingual scripts, and optimizing it for deployment on resource-constrained devices. Ultimately, this research highlights the enduring value of high-quality input preparation in achieving reliable, accurate, and inclusive handwriting recognition.

Author Contributions: All authors contributed to the research and manuscript preparation equally.

Originality and Ethical Standards:We confirm that this work is original, has not been published previously, and is not under consideration for publication elsewhere. All ethical standards, including proper citations and acknowledgments, have been strictly followed in the preparation of this manuscript.

Data availability: Data available upon request.

Conflict of Interest: There is no conflict of Interest.

Funding: The research received no external funding.

1. N. Otsu, "A threshold selection method from gray-level histograms," IEEE Transactions on Systems, Man, and Cybernetics, vol. 9, no. 1, pp. 62–66, 1979.
2. A. Graves, M. Liwicki, S. Fernández, R. Bertolami, H. Bunke, and J. Schmidhuber, "A novel connectionist system for unconstrained handwriting recognition," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 31, no. 5, pp. 855–868, May 2009.
3. P. Y. Simard, D. Steinkraus, and J. C. Platt, "Best practices for convolutional neural networks applied to visual document analysis," in Proc. Int. Conf. Document Analysis and Recognition (ICDAR), 2003, pp. 958–963.
4. S. Hochreiter and J. Schmidhuber, "Long short-term memory," Neural Computation, vol. 9, no. 8, pp. 1735–

1780, 1997.

1. A. Vaswani et al., "Attention is all you need," in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2017, pp. 5998–6008.
2. T. Bluche, "Joint line segmentation and transcription for end-to-end handwritten paragraph recognition," in Advances in Neural Information Processing Systems, vol. 28, 2015.
3. C. Tensmeyer and T. Martinez, "Document image binarization with fully convolutional neural networks," in Proc. Int. Conf. Document Analysis and Recognition (ICDAR), 2017, pp. 99–104.
4. M. Pham, T. Bluche, C. Kermorvant, and J. Louradour, "Dropout improves recurrent neural networks for handwriting recognition," in Proc. Int. Conf. Frontiers in Handwriting Recognition (ICFHR), 2014, pp. 285–290.
5. M. Abdelaziz, M. Elaraby, and S. A. Mahmoud, "Handwriting recognition using deep learning: A comprehensive review," Journal of Imaging, vol. 6, no. 8,

p. 73, Aug. 2020.

1. T. K. Bhowmik and P. P. Roy, "Deep learning approaches for handwriting recognition: A review," Artificial Intelligence Review, vol. 53, pp. 5367–5406, 2020.
2. M. Liwicki and H. Bunke, "IAM-OnDB—An online English sentence database acquired from handwritten text on a whiteboard," in Proc. Int. Conf. Document Analysis and Recognition (ICDAR), 2005, pp. 956–961.
3. U. V. Marti and H. Bunke, "The IAM-database: An English sentence database for offline handwriting recognition," International Journal on Document Analysis and Recognition, vol. 5, no. 1, pp. 39–46, 2002.
4. D. Doermann, "The role of document image analysis in understanding digital content," Pattern Recognition Letters, vol. 35, pp. 2–11, Jan. 2014.
5. C. L. Liu, H. Koga, and H. Fujisawa, "Lexicon-driven segmentation and recognition of handwritten character strings for OCR," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 24, no. 11, pp. 1425–1437,

Nov. 2004.

1. R. Plamondon and S. N. Srihari, "On-line and off-line handwriting recognition: A comprehensive survey," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 22, no. 1, pp. 63–84, Jan. 2000.
2. J. Wolf, J.-M. Jolion, and F. Chassaing, "Text localization, enhancement and binarization in multimedia documents," in IEEE Workshop on Multimedia Signal Processing, 2002, pp. 223–228.
3. G. Louloudis, B. Gatos, and I. Pratikakis, "Text line detection in handwritten documents," Pattern Recognition, vol. 41, no. 12, pp. 3758–3772, Dec. 2008.
4. K. Raza and I. Siddiqi, "An offline cursive script recognition system for Urdu," Pattern Recognition, vol. 48, no. 5, pp. 1730–1742, May 2015.