DeQA-Doc: Adapting DeQA-Score to Document Image Quality Assessment

Junjie Gao, Runze Liu, Yingzhe Peng, Shujian Yang, Jin Zhang, Kai Yang, Zhiyuan YouarXiv 2025 · DOI · OpenReview

Abstract

Document quality assessment is critical for a wide range of applications including document digitization, OCR, and archival. However, existing approaches often struggle to provide accurate and robust quality scores, limiting their applicability in practical scenarios. With the rapid progress in Multi-modal Large Language Models (MLLMs), recent MLLM-based methods have achieved remarkable performance in image quality assessment. In this work, we extend this success to the document domain by adapting DeQA-Score, a state-of-the-art MLLM-based image quality scorer, for document quality assessment. We propose DeQA-Doc, a framework that leverages the visual language capabilities of MLLMs and a soft label strategy to regress continuous document quality scores. To adapt DeQA-Score to DeQA-Doc, we adopt two complementary solutions to construct soft labels without the variance information. Also, we relax the resolution constrains to support the large resolution of document images. Finally, we introduce ensemble methods to further enhance the performance. Extensive experiments demonstrate that DeQA-Doc significantly outperforms existing baselines, offering accurate and generalizable document quality assessment across diverse degradation types.

Introduction

Assessing the quality of document images—whether scanned or photographed—presents unique challenges compared to general natural images. Document Image Quality Assessment (DIQA) requires sensitivity to fine-grained text details and complex layouts. Traditional methods often rely on hand-crafted features that fail to capture high-level semantics, while standard deep learning models frequently struggle to generalize across varied degradations like blur, noise, or uneven lighting.

Overview of the DeQA-Score framework mapping continuous scores to discrete level tokens.

Recent advances in Multi-modal Large Language Models (MLLMs) have shown promise in general Image Quality Assessment (IQA). However, adapting these models to the document domain faces three primary hurdles: the lack of data variance in document datasets (which typically only provide a Mean Opinion Score), resolution bottlenecks in vision encoders (like CLIP) that "smear" fine text, and sensitivity to prompt phrasing.

Methods

DeQA-Doc addresses these challenges by extending the DeQA-Score framework with several key adaptations.

Soft Label Construction

To enable the model to output continuous scores, DeQA-Doc employs a distribution-based strategy. Since document datasets often lack variance data, two workarounds were proposed:

  1. Pseudo-variance: A fixed variance (approximately 20% of the score range) is assigned based on empirical statistics from general IQA datasets like KonIQ-10K and SPAQ.
  2. Linear Interpolation: Probability mass is distributed between the two nearest discrete quality levels (e.g., "fair" and "good").

Resolution and Architecture

To preserve text legibility, the framework relaxes resolution constraints. For the mPLUG-Owl2 backbone, absolute position embeddings were removed to support arbitrary image sizes. The model also integrates Qwen2.5-VL, which natively supports dynamic resolutions.

The DeQA-Doc architecture utilizing vision and text tokens with a hybrid loss function.

Training and Inference

The model is trained using a hybrid objective: standard Cross-Entropy loss for text responses and KL Divergence loss to align predicted quality distributions with soft labels. To enhance robustness, DeQA-Doc utilizes model ensembles (averaging mPLUG-Owl2 and Qwen2.5-VL) and prompt ensembles (averaging results from multiple query variations).

Results and Discussion

Evaluated on the DIQA-5000 benchmark, DeQA-Doc achieved a state-of-the-art score of 0.9288, significantly outperforming the previous best baseline, Q-Align (0.8523).

Key findings include:

  • Resolution Sweet Spot: Increasing resolution to 1024x1024 provided a substantial performance boost. However, further increases to 1536x1536 led to a slight decline, suggesting a point of diminishing returns where noise or redundant details may interfere.
  • Effectiveness of Pseudo-variance: Modeling quality as a Gaussian distribution using pseudo-variance proved more effective than simple linear interpolation, highlighting the importance of capturing perceptual uncertainty.
  • Architecture Synergy: Ensembling different MLLM backbones yielded better results than single-model approaches, indicating that different architectures possess complementary strengths in document feature perception.

Conclusion

DeQA-Doc successfully bridges the gap between general image perception and specialized document analysis. By introducing pseudo-variance for soft-label training and optimizing for higher resolutions, the framework sets a new standard for DIQA. Future work may explore unifying image and document quality assessment within a single multi-modal framework to further enhance generalizability.

Summary updated

Abstract

Document quality assessment is critical for a wide range of applications including document digitization, OCR, and archival. However, existing approaches often struggle to provide accurate and robust quality scores, limiting their applicability in practical scenarios. With the rapid progress in Multi-modal Large Language Models (MLLMs), recent MLLM-based methods have achieved remarkable performance in image quality assessment. In this work, we extend this success to the document domain by adapting DeQA-Score, a state-of-the-art MLLM-based image quality scorer, for document quality assessment. We propose DeQA-Doc, a framework that leverages the visual language capabilities of MLLMs and a soft label strategy to regress continuous document quality scores. To adapt DeQA-Score to DeQA-Doc, we adopt two complementary solutions to construct soft labels without the variance information. Also, we relax the resolution constrains to support the large resolution of document images. Finally, we introduce ensemble methods to further enhance the performance. Extensive experiments demonstrate that DeQA-Doc significantly outperforms existing baselines, offering accurate and generalizable document quality assessment across diverse degradation types. Codes and model weights are available in https://github.com/Junjie-Gao19/DeQA-Doc.
BibTeX
@misc{gao2025deqa,
  author = {Junjie Gao and Runze Liu and Yingzhe Peng and Shujian Yang and Jin Zhang and Kai Yang and Zhiyuan You},
  title = {{DeQA-Doc: Adapting DeQA-Score to Document Image Quality Assessment}},
  year = {2025},
  eprint = {2507.12796},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.12796},
}
Figure