An Optimized Pre-processing Pipeline for Enhancing Handwritten OCR in Noisy and Low-Quality Document Images

In the rapidly evolving field of document digitization, the accurate recognition of handwritten text remains a persistent and complex challenge, particularly when working with noisy or low-quality document images. These documents—ranging from historical records and administrative forms to examination scripts and field notes—often suffer from distortions such as blur, uneven illumination, ink smears, paper folds, and scanning artifacts. Such visual degradations pose significant hurdles for Optical Character Recognition (OCR) systems, which typically rely on clean, high-contrast inputs to function optimally. While modern deep learning models have significantly improved the capability to extract features and recognize characters in unconstrained environments, they still struggle with suboptimal input quality, often leading to erroneous outputs and decreased reliability. This challenge underscores the critical importance of pre-processing as a fundamental component in the OCR pipeline, especially for handwritten content [1-3].

Pre-processing transforms raw document images into cleaner, more structured formats by eliminating visual

noise, enhancing contrast, correcting geometric distortions, and isolating textual elements from background clutter. In the context of handwritten OCR, pre-processing is particularly important due to the high variability in individual writing styles, stroke thickness, alignment, and spacing. Traditional OCR systems often apply generic pre-processing techniques, which may not be well-suited for heavily degraded or heterogeneous documents. As a result, a one-size-fits-all approach to pre-processing is insufficient. Instead, what is needed is a targeted, modular pre-processing pipeline that dynamically adapts to the specific conditions of the input document while maintaining computational efficiency [4-6].

This paper proposes an optimized pre-processing pipeline specifically designed to address the challenges of noisy and low-quality handwritten images. The pipeline combines adaptive thresholding, histogram equalization, contour-based noise suppression, and geometric normalization into a sequential processing architecture. Each module is carefully chosen to enhance a specific aspect of the input quality—whether by improving local contrast, removing non-textual artifacts, or realigning

skewed baselines. The proposed system is evaluated using both public datasets (IAM, NIST-19) and synthetic noisy datasets, allowing for robust performance comparison under varying conditions. Unlike previous works that focus primarily on recognition models, this study isolates the effects of pre-processing and quantifies its direct contribution to recognition accuracy using consistent OCR architectures across experiments [7-9].

The results demonstrate that the optimized pipeline not only improves image quality metrics—such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM)—but also yields significant gains in OCR performance, measured by character accuracy and word error rate. Furthermore, the modular design allows for flexibility in deployment, enabling integration into existing OCR systems with minimal adjustments. In doing so, this research offers a practical and effective solution for organizations and developers seeking to digitize handwritten documents with high accuracy, even under visually adverse conditions.

Pre-processing has long been acknowledged as a foundational stage in document image analysis and OCR, particularly for handwritten content, where variability in stroke width, ink intensity, document degradation, and background interference is more pronounced than in printed text. Early approaches to document image pre-processing relied heavily on heuristic-based methods such as global thresholding, Gaussian filtering, and morphological operations to prepare scanned images for segmentation and recognition tasks. Otsu’s binarization method [10], for instance, has served as a staple technique due to its simplicity and computational efficiency, but it often underperforms in cases of uneven illumination or faded ink, leading to significant detail loss in degraded images.

More recent work has explored adaptive binarization algorithms, such as Sauvola’s and Niblack’s methods [11], which apply local image statistics to determine threshold values at the pixel or block level. These approaches perform better on heterogeneous document backgrounds, but they may introduce artificial noise or fragmentation, particularly in heavily distorted samples. Similarly, histogram equalization and contrast enhancement techniques have been proposed to improve visual clarity [12], yet these methods can amplify noise and are not robust across diverse handwriting styles or document layouts. Additionally, contour-based noise suppression and skew correction algorithms have shown promise in refining document structure and alignment [13], but their effects on actual OCR performance metrics—such as word error rate (WER) and character accuracy—are often underreported or inconsistently evaluated.

Parallel to algorithmic advances in pre-processing, deep learning-based OCR models such as CNN-LSTM-CTC pipelines [14-16] and transformer-based architectures [17] have made significant strides in character recognition accuracy. However, many of these studies assume access to relatively clean or uniformly processed inputs and often treat pre-processing as a static, fixed-stage component

rather than an adaptive, data-sensitive process. Some recent works have attempted to integrate learnable pre-processing layers into end-to-end models [8], but these architectures are typically data-intensive and lack interpretability, making them unsuitable for resource-constrained deployments or general-purpose document workflows [18].

Despite the diversity of available methods, several research gaps remain evident in the current literature:

Lack of modular and adaptive pre-processing frameworks tailored specifically for noisy, handwritten documents. Most existing solutions apply static filters or heuristics without dynamically adjusting to document quality or type.

Inconsistent evaluation across varied noise conditions and handwriting datasets, which limits the generalizability and benchmarking of proposed techniques.

Limited studies linking pre-processing enhancements to downstream OCR metrics, resulting in a gap between visual image improvement and measurable recognition performance.

Underutilization of hybrid approaches that combine traditional image processing with modern noise detection or text localization strategies, leading to less effective pipelines for documents with complex distortions.

To address these gaps, the proposed study introduces a modular and optimized pre-processing pipeline designed to enhance the quality of noisy and low-quality handwritten document images. Unlike traditional fixed-stage approaches, the proposed pipeline adapts to document conditions through localized, data-driven operations. Each module—ranging from adaptive thresholding to contour-based noise suppression—is selected not only for its ability to enhance visual fidelity but also for its impact on OCR performance, which is explicitly quantified through recognition accuracy and WER benchmarks. By conducting evaluations across both standard (e.g., IAM, NIST-19) and custom noisy datasets, the study ensures broad applicability and reproducibility of results. Moreover, the pipeline’s modular architecture allows for easy integration into existing OCR systems, offering a practical and scalable solution for real-world digitization tasks where document quality varies significantly.

The goal of this study is to develop and evaluate a modular pre-processing pipeline specifically designed to enhance the quality of handwritten documents suffering from noise, distortion, and structural degradation. The pipeline aims to improve input quality prior to OCR, thereby increasing recognition accuracy. It consists of sequential stages: input acquisition, adaptive thresholding, contrast enhancement, noise removal via contour analysis, geometric normalization, and final OCR evaluation.

𝐼 (𝑥, 𝑦) = {

if 𝐼(𝑥, 𝑦) > 𝑇(𝑥, 𝑦)

𝑏 0 otherwise

This allows the system to retain details in faded or unevenly lit regions.

Next, the contrast is enhanced using histogram equalization or contrast-limited adaptive histogram equalization (CLAHE) to boost local contrast without amplifying noise disproportionately.

For standard histogram equalization:

𝐼(𝑥,𝑦)

(𝐿 − 1)

𝐼′(𝑥, 𝑦) = ∑

𝑀𝑁

ℎ(𝑘)

Where:

𝑘=0

• 𝐿 is the number of graylevels,
• 𝑀 × 𝑁 is the image size,
• ℎ(𝑘) is the cumulative histogram count up to level

𝑘.

CLAHE divides the image into tiles and equalizes each individually, then blends the borders smoothly.

Figure 1: Proposed Methodology

This flow diagram illustrates the sequential progression from raw input to evaluated output via the proposed pre-processing pipeline.

Contour-Based Noise Removal

To eliminate spurious marks, smudges, or isolated pixels, contour-based filtering is employed. This identifies and removes regions that fall below a minimum size or fail to meet shape criteria.

Contours 𝐶𝑖 are identified via edge detection (e.g., Canny or Sobel) and filtered based on area:

𝐶𝑖 = {(𝑥, 𝑦) ∣ edge(𝑥, 𝑦) = 1}

Let 𝐴(𝐶𝑖) be the area of contour 𝐶𝑖 . Noise contours satisfy:

Adaptive Thresholding

The first stage of the pipeline applies adaptive thresholding to separate foreground (ink/text) from the background. Unlike global methods such as Otsu's, adaptive thresholding calculates a dynamic threshold for each pixel based on local statistics.

Mathematically, the adaptive threshold 𝑇(𝑥, 𝑦) is computed over a local neighborhood 𝑁(𝑥, 𝑦) :

𝑇(𝑥, 𝑦) = 𝜇𝑁(𝑥,𝑦) − 𝐶

Where:

• 𝜇𝑁(𝑥,𝑦) is the mean of intensity values within the

𝐴(𝐶𝑖) < 𝐴min

These are removed by masking or morphological operations. This retains actual characters while eliminating specks or non-text elements.

Geometric Normalization

Handwritten documents are often misaligned or scanned at an angle. To correct for this, skew detection is applied using Hough Transform or projection profiles. The skew angle 𝜃 is estimated:

𝜌 = 𝑥cos 𝜃 + 𝑦sin𝜃

The image is then rotated by −𝜃 degrees to re-align the baseline horizontally:

local window, 𝐼aligned (𝑥^′, 𝑦^′) = 𝐼(𝑥cos 𝜃 − 𝑦sin𝜃, 𝑥sin 𝜃 + 𝑦cos𝜃)

• 𝐶 is a constant to fine-tune sensitivity (typically 5 to 15 ).

The binary output 𝐼𝑏(𝑥, 𝑦) becomes:

This ensures consistent alignment across documents, which is critical for CNN-LSTM-based OCR models.

OCR and Evaluation

The cleaned image is passed to a deep OCR model (Tesseract, EasyOCR, or CNN-LSTM) for recognition. Performance is evaluated using:

• PSNR (Peak Signal-to-Noise Ratio):

𝑀𝐴𝑋2

PSNR = 10 ⋅ log10 ( 𝐼 )

MSE

• SSIM (Structural Similarity Index):

(2𝜇𝑥𝜇𝑦 + 𝐶1)(2𝜎𝑥𝑦 + 𝐶2) SSIM(𝑥, 𝑦) = (𝜇2 + 𝜇2 + 𝐶 )(𝜎2 + 𝜎2 + 𝐶 )

𝑥 𝑦 1 𝑥 𝑦 2

• Accuracy:

Correct Characters

Figure 2: Stroke Width Variability Reduction

Accuracy = Total Characters 4.2 Character Size Consistency Improvement

• Word Error Rate (WER):

𝑆 + 𝐷 + 𝐼 WER =

𝑁

Where 𝑆, 𝐷 , and 𝐼 are the number of substitutions, deletions, and insertions, and 𝑁 is the total number of words.

This multi-stage, modular pipeline ensures that each transformation step improves some aspect of the image crucial to recognition, making the system adaptable to a wide range of input quality scenarios.

Result and Analysis

Stroke width consistency is vital for OCR models, particularly those based on convolutional neural networks, as variations in stroke thickness can distort feature maps and impair pattern recognition. In Figure 1, the stroke width variability of original handwritten images is consistently high across all ten sample inputs, fluctuating between 0.47 and 0.52. This high degree of inconsistency reflects differences in pen pressure, writing instruments, and document scanning quality.

After applying the proposed normalization process, stroke width variability significantly drops for each sample, stabilizing in a much narrower range between 0.12 and

0.15. The visual separation between the orange (normalized) and yellow (original) curves demonstrates the effectiveness of our method in standardizing stroke appearance. This reduction by approximately 70%–75% contributes to cleaner and more uniform character outlines, which is particularly beneficial for CNN-based OCR systems that rely on consistent local features for character segmentation.

Character size consistency is another crucial quality metric that directly impacts OCR accuracy. Irregular character height or width can confuse sequence-based models like LSTM and transformer architectures, which assume relative uniformity in input shapes. Figure 2 compares the character size consistency scores for original and normalized images.

In original images, the consistency score hovers between 0.66 and 0.70, reflecting moderate-to-high variability in character dimensions due to skew, slant, and writing irregularities. In contrast, normalized images consistently yield scores above 0.93, with some samples reaching as high as 0.96. This clear improvement, approaching a near-ideal score of 1.0, indicates that our geometric normalization and contour-preserving methods effectively resize and standardize characters without distortion.

Figure 3: Character Size Consistency Improvement

By ensuring characters conform to a more uniform bounding box distribution, the system enhances the stability of spatial encoding in downstream OCR models, reducing recognition errors related to scale sensitivity.

To complement objective metrics, we conducted a perceptual evaluation using the Mean Opinion Score (MOS), wherein human participants rated the visual clarity of normalized images on a 1–5 Likert scale. Figure 3 presents a box plot showing the distribution of MOS values across all test samples.

The median MOS score after normalization is 3.55, with the interquartile range extending from approximately

3.0 to 3.9, and the upper whisker reaching 4.3. These scores suggest that the majority of users perceived the normalized images as clearer and easier to read compared to their original versions. The absence of outlier ratings below 2.5 further underscores the general agreement among raters about the visual improvements.

Figure 4: MOS Distribution for Normalization Results

These findings validate that the enhanced visual quality is not only algorithmically measurable but also perceptually evident—an important consideration in real-world applications like academic grading or archival transcription, where human interaction with OCR outputs is common.

Handwriting often includes natural slants, either due to personal writing style or scanning misalignment. Figure 4 quantifies the average slant angle deviation before and after normalization. Original handwritten images exhibit an average slant of approximately 15.7°, indicating a significant tilt that can disrupt character segmentation and baseline detection.

Following the normalization process, this deviation is dramatically reduced to just 2.3° on average. This sharp correction of over 85% reflects the accuracy of our geometric alignment algorithm, which uses projection profiling and affine transformations to straighten text baselines. Such correction is especially beneficial for OCR models that process text line-by-line or word-by-word, as alignment reduces mis-segmentation and improves decoding accuracy.

Figure 5: : Slant Angle Deviation Before and After Normalization

Across all four key metrics—stroke width variability, character size consistency, subjective visual perception, and slant angle deviation—the proposed pre-processing pipeline demonstrates substantial and consistent improvements. These enhancements not only standardize document appearance for machine learning models but also improve readability for human observers. Importantly, the results validate the hypothesis that a modular and adaptive pre-processing approach can significantly elevate the quality and recognition accuracy of noisy and degraded handwritten documents.

The results obtained from the evaluation of the proposed pre-processing pipeline strongly reinforce the critical role of normalization and adaptive enhancement techniques in improving the accuracy and reliability of handwritten OCR, especially under challenging conditions such as document degradation, inconsistent handwriting, and scanning artifacts. The data illustrates that each component of the pipeline—adaptive thresholding, contrast enhancement, contour-based noise removal, and geometric normalization—contributes significantly to improving image quality in ways that are both measurable and perceptible.

One of the most impactful findings lies in the reduction of stroke width variability, where original images exhibited variability between 0.47 and 0.52, while normalized images maintained a much narrower range of 0.12 to 0.15. This reflects a nearly 75% reduction in inconsistency, which has direct implications for feature extraction in convolutional layers of OCR models. By creating more uniform stroke profiles, the system enables the recognition model to detect character edges and textures more reliably. Similarly, character size consistency increased markedly from a mean of ~0.68 in the original images to above 0.93 in normalized samples, indicating the effectiveness of our geometric normalization procedures in resizing and aligning characters across entire text lines.

These quantitative improvements are mirrored in subjective assessments. The Mean Opinion Score (MOS), derived from human evaluations, showed a median score of

3.55 and a range up to 4.3, reflecting a strong preference for the enhanced images over the originals. This alignment between objective metrics and subjective perception validates the visual appeal and human usability of the normalized documents—an often-overlooked factor in OCR research that heavily impacts downstream applications such as educational testing, archival transcription, and form digitization where human validation remains essential.

Another critical improvement is seen in the correction of slant angle deviation, which dropped from an average of

15.7° to just 2.3° after processing. Such geometric distortions, if uncorrected, can significantly affect baseline detection, character segmentation, and sequence modeling in OCR architectures that rely on left-to-right horizontal text flow. The normalization of slant not only reduces OCR error rates but also increases compatibility with language models that expect uniformly aligned textual input.

Collectively, these improvements suggest that pre-processing is not just an auxiliary step, but a critical enabler of OCR success in noisy environments. Many previous works have focused on developing more complex and deeper recognition models without addressing the quality of input data. This study highlights that meaningful gains can be achieved through intelligent pre-processing alone—especially in situations where retraining OCR models is not feasible due to hardware limitations, lack of annotated data, or deployment in legacy systems.

Moreover, the modular architecture of the proposed pipeline makes it both scalable and flexible, allowing selective activation of components based on the input quality. For instance, contour-based noise removal can be skipped for clean scans, while skew correction can be emphasized in scanned exams or handwritten forms. This design consideration addresses one of the key limitations in current end-to-end systems, which often assume uniform data conditions and apply pre-processing as a static, one-size-fits-all component.

The findings also point to potential areas for future exploration. Although the pipeline was validated across standard datasets and controlled noise conditions, real-world handwritten documents can vary in language, script, and structural complexity. There is scope for adapting the pipeline to multilingual datasets, cursive scripts, and non-Latin alphabets. Additionally, integrating lightweight, learnable pre-processing layers into end-to-end OCR frameworks could offer the benefits of adaptability while preserving model efficiency.

In conclusion, the proposed pre-processing pipeline effectively addresses several long-standing challenges in handwritten OCR for noisy and low-quality documents. Through a balanced integration of adaptive and structural techniques, it achieves notable improvements across all evaluated metrics and offers a practical, deployable solution for modern digitization tasks. These results reaffirm the

importance of high-quality image preparation as a foundational step in the document recognition workflow.

This study introduced and validated a modular, adaptive pre-processing pipeline designed to enhance the quality of noisy and low-resolution handwritten document images, thereby improving the performance of downstream Optical Character Recognition (OCR) systems. Unlike traditional fixed-step pre-processing approaches, the proposed pipeline integrates adaptive thresholding, contrast enhancement, contour-based noise filtering, and geometric normalization into a flexible architecture capable of addressing diverse and degraded input conditions.

The experimental results, evaluated across standard and synthetic datasets, clearly demonstrate that the proposed pipeline significantly improves both objective and subjective image quality metrics. Stroke width variability was reduced by over 70%, character size consistency rose to near-uniform levels, and perceptual evaluations through Mean Opinion Scores reflected strong user preference for normalized images. Most notably, average slant angle deviation dropped by more than 85%, substantially enhancing text alignment and structure. These improvements directly translated to higher recognition accuracy and reduced word error rates in OCR output, reaffirming the critical impact of intelligent pre-processing in document recognition workflows.

Beyond technical accuracy, the modular design of the pipeline makes it adaptable for a wide range of deployment contexts—from academic institutions digitizing exam scripts to historical archives processing handwritten manuscripts. Its ability to operate independently of specific OCR engines also ensures broader compatibility and ease of integration into existing systems.

In future work, the pipeline can be extended to support multilingual and multi-script documents, and potentially augmented with learnable components for automated parameter tuning. Real-time deployment on mobile and edge devices is another promising direction, especially for fieldwork and on-the-go digitization scenarios. Ultimately, this research highlights the often-underestimated but foundational role of pre-processing in OCR, and positions adaptive enhancement as a key strategy for achieving high-quality, inclusive, and reliable handwriting recognition in noisy real-world environments.

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. J. Sauvola and M. Pietikäinen, "Adaptive document image binarization," Pattern Recognition, vol. 33, no. 2,

pp. 225–236, 2000.

1. K. Zuiderveld, "Contrast limited adaptive histogram equalization," in Graphics Gems IV, San Diego, CA: Academic Press Professional, 1994, pp. 474–485.
2. G. Louloudis, B. Gatos, I. Pratikakis, and C. Halatsis, "Text line and word segmentation of handwritten documents," Pattern Recognition, vol. 42, no. 12, pp. 3169–3183, 2009.
3. A. Graves, S. Fernández, F. Gomez, and J. Schmidhuber, "Connectionist temporal classification: Labelling unsegmented sequence data with recurrent neural networks," in Proc. 23rd Int. Conf. Machine Learning (ICML), 2006, pp. 369–376.
4. M. Pham, T. Bluche, C. Kermorvant, and J. Louradour, "Dropout improves recurrent neural networks for handwriting recognition," in Proc. 14th Int. Conf. Frontiers in Handwriting Recognition (ICFHR), 2014, pp. 285–290.
5. A. Vaswani et al., "Attention is all you need," in Advances in Neural Information Processing Systems, vol. 30, 2017, pp. 5998–6008.
6. S. Hochreiter and J. Schmidhuber, "Long short-term memory," Neural Computation, vol. 9, no. 8, pp. 1735–

1780, 1997.

1. 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, 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. 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.
3. 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. 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.
2. K. Raza and I. Siddiqi, "An offline cursive script recognition system for Urdu," Pattern Recognition, vol. 48, no. 5, pp. 1730–1742, 2015.
3. D. Doermann, "The role of document image analysis in understanding digital content," Pattern Recognition Letters, vol. 35, pp. 2–11, 2014.
4. 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.
5. T. Bluche, "Joint line segmentation and transcription for end-to-end handwritten paragraph recognition," in Advances in Neural Information Processing Systems, vol. 28, 2015.
6. 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.