DeQA-Doc: Document Image Quality Assessment

• DeQA-Doc is a document image quality assessment framework that formulates DIQA as continuous score regression using a soft distribution over five pre-defined quality levels.
• It employs pseudo-variance and linear interpolation soft-label constructions along with resolution-flexible vision encoding to capture text legibility and layout integrity in high-resolution document images.
• The framework achieves robustness through model ensembling and demonstrates superior performance on DIQA-5000, with a final score reaching 0.9288.

DeQA-Doc is a framework for Document Image Quality Assessment (DIQA) that adapts DeQA-Score, a Multi-modal LLM (MLLM)-based image quality scorer, to the document domain. It formulates DIQA as continuous score regression from a document image to a scalar quality value, but does so indirectly: the model predicts a soft distribution over five text-defined quality levels—bad, poor, fair, good, and excellent—and recovers a continuous score as the expectation of that distribution. The framework is motivated by document-processing settings in which image quality directly affects OCR, digitization workflows, archival systems, and broader document understanding pipelines (Gao et al., 17 Jul 2025).

1. Problem setting and conceptual basis

In DeQA-Doc, DIQA is treated as a problem of estimating how readable, usable, or visually acceptable a document image is. The paper emphasizes that this is not merely a low-level distortion-detection task. For documents, quality is tied to text legibility, layout preservation, and whether degradations impair downstream use. The motivating degradation patterns include blur, low contrast, shadows, moiré, compression artifacts, uneven lighting, and related distortions that can reduce recognition accuracy or distort document structure (Gao et al., 17 Jul 2025).

The framework is explicitly positioned against two prior classes of DIQA methods. First, traditional handcrafted-feature or shallow-model methods are said to rely on manually designed signals such as gradients or image statistics and to “often fail to capture high-level semantic and structural information.” Second, more recent deep-learning-based methods improve over handcrafted methods but are argued to “struggle to generalize across diverse document layouts and degradation types.” DeQA-Doc therefore adopts an MLLM-based formulation because document quality is presented as a semantically structured visual reasoning problem rather than only a generic image-regression task (Gao et al., 17 Jul 2025).

A central conceptual inheritance comes from DeQA-Score, originally developed for natural-image quality assessment. DeQA-Doc retains the idea of replacing direct scalar regression with prediction of a soft distribution over discrete quality levels. The model is trained to emit a templated response such as “The quality of this image is <level>”, and the probability assigned to the quality token becomes the basis for the final continuous score. This means that DeQA-Doc is best understood not as a wholly new DIQA formulation, but as a document-specific extension of DeQA-Score to high-resolution, text-centric, structurally constrained images (Gao et al., 17 Jul 2025).

2. Score-distribution formulation and soft-label construction

The core scoring mechanism uses five ordered quality levels with numeric centers

$c_i \in \{1,2,3,4,5\}.$

If $p_i^{pred}$ denotes the model’s predicted probability for level $i$, the continuous prediction is

$\mu^{pred} = \sum\nolimits_i p_i^{pred} c_i.$

This is the central regression mechanism in DeQA-Doc: the model does not attach a separate regression head to the backbone, but instead derives a scalar score from a language-aligned categorical distribution (Gao et al., 17 Jul 2025).

The main technical obstacle in transferring DeQA-Score to documents is that DIQA-5000 provides only MOS, not variance. DeQA-Score’s original soft-label construction assumes a Gaussian quality distribution

$x \sim \mathcal{N}(\mu,\sigma^2),$

discretized into five bins centered at $c_i$, with raw bin probabilities

$$p_i^{raw} = \int_{c_i - d/2}^{c_i + d/2}f(x)\,{\rm dx},$$

followed by an adjustment so that the recovered mean matches the MOS. Because DIQA-5000 lacks $\sigma^2$, this procedure cannot be applied directly in the document setting (Gao et al., 17 Jul 2025).

DeQA-Doc therefore introduces two complementary soft-label constructions without variance information. The first is pseudo variance. It draws on an empirical observation from DeQA-Score that, on IQA datasets such as KonIQ, SPAQ, and KADID, the average standard deviation is roughly 20% of the full score range. DeQA-Doc thus sets the pseudo standard deviation to

$0.2 \times (\text{max} - \text{min}),$

normalizes scores to $[1,5]$, and then reuses the Gaussian-to-discrete labeling pipeline. The second is linear interpolation between adjacent quality levels. For a mean score $p_i^{pred}$0 between adjacent centers $p_i^{pred}$1 and $p_i^{pred}$2, the target mass is assigned only to those two bins, yielding a sparse two-level target distribution (Gao et al., 17 Jul 2025).

Training combines ordinary autoregressive cross-entropy over the response template with a KL divergence on the quality-level token. The paper’s printed KL formula is malformed, but its intended role is clear: the predicted level distribution is matched against the soft target distribution. The result is a hybrid training objective in which ordinary language modeling constrains the response form, while the quality token receives explicit soft-label supervision (Gao et al., 17 Jul 2025).

3. Document-specific adaptations

The paper identifies three document-specific adaptations required to make DeQA-Score work well on DIQA.

The first is precisely the soft-label redesign without variance annotations. This is not a minor implementation detail. In the document domain, the lack of variance in DIQA-5000 makes the original DeQA-Score label construction infeasible, so DeQA-Doc’s pseudo-variance and linear-interpolation schemes are foundational rather than optional. Empirically, the paper prefers pseudo variance, which suggests that a broader target distribution better captures perceptual uncertainty than the limiting two-bin interpolation case (Gao et al., 17 Jul 2025).

The second is high-resolution document support. DeQA-Score’s original mPLUG-Owl2 configuration assumes fixed image resolution, typically $p_i^{pred}$3, which is problematic for documents because aggressive downsampling can destroy small text, layout boundaries, and subtle quality cues. DeQA-Doc addresses this in two ways. One option is to modify mPLUG-Owl2-7B by removing the absolute position embeddings from the CLIP-based vision encoder, thereby relaxing the fixed-resolution constraint. The other is to use Qwen2.5-VL-7B, which natively accepts dynamic and original resolutions. The paper is explicit that this adaptation is achieved through resolution-flexible vision encoding, not through tiling, multi-scale fusion, sliding windows, OCR extraction, or explicit cropping pipelines (Gao et al., 17 Jul 2025).

The third is ensemble design for robustness. DeQA-Doc studies both model ensemble and prompt ensemble. Model ensemble averages the predicted five-way score distributions from multiple models element-wise before taking the expectation-based score. Prompt ensemble averages predictions over 10 semantically equivalent prompts. The empirical outcome is asymmetric: model ensemble yields substantial gains, whereas prompt ensemble contributes little. This suggests that, in this task, the dominant uncertainty lies in model variance rather than prompt phrasing (Gao et al., 17 Jul 2025).

4. Architecture, training procedure, and inference

DeQA-Doc is instantiated with two MLLM families: mPLUG-Owl2-7B and Qwen2.5-VL-7B. The architecture follows a standard MLLM decomposition into vision encoder, vision-to-text connector, and LLM. The model receives a document image, encodes it visually, projects visual tokens into the language embedding space, and then generates a templated textual answer whose critical token is one of the five quality levels (Gao et al., 17 Jul 2025).

The principal dataset is DIQA-5000. The paper uses its training split for optimization, its validation split for ablations and development, and its hidden test split for final evaluation. The labels cover overall quality, sharpness, and color fidelity. The official benchmark score combines PLCC and SRCC for these dimensions into a weighted final score, with overall quality weighted more heavily than sharpness and color fidelity; the printed equations in the paper are corrupted, but the weighting structure is described clearly (Gao et al., 17 Jul 2025).

Training explores both full parameter tuning and LoRA fine-tuning. The reported optimization setup uses AdamW, an initial learning rate of $p_i^{pred}$4, cosine decay, and 3 epochs, on 8 NVIDIA RTX A100 GPUs. For mPLUG-Owl2-based runs, the paper studies input resolutions of $p_i^{pred}$5, $p_i^{pred}$6, and $p_i^{pred}$7; for Qwen2.5-VL, it also uses original resolution. Batch sizes are 64 at $p_i^{pred}$8, 8 at $p_i^{pred}$9, and 2 at $i$0. The paper also studies optional pretraining on KonIQ, a natural-image IQA dataset, before adapting to DIQA-5000 (Gao et al., 17 Jul 2025).

At inference time, the model extracts the predicted probabilities for the five level tokens, computes the expected score $i$1, and, if desired, averages score distributions across prompts and/or models. The response template remains tightly constrained; the task is not open-ended captioning but a controlled quality-assessment response whose semantics are then mapped back into a continuous DIQA score (Gao et al., 17 Jul 2025).

5. Empirical performance and ablation results

The paper reports a sequence of ablations that isolate the contribution of each design choice. On DIQA-5000 validation, Q-Align attains a final score of 0.8523, whereas DeQA-Score (w/o fidelity loss) reaches 0.8630, and DeQA-Score (with fidelity loss) reaches 0.8849. This establishes that DeQA-Doc benefits from the score-distribution formulation rather than from generic MLLM prompting alone (Gao et al., 17 Jul 2025).

Resolution matters substantially. Validation performance improves from 0.8849 at $i$2 to 0.8989 at $i$3, but drops to 0.8861 at $i$4. The paper interprets this as indicating that higher resolution is beneficial up to a point, after which redundant detail or overfitting may become harmful. The soft-label ablation shows pseudo variance outperforming linear interpolation, with final scores of 0.8989 versus 0.8714. The fine-tuning ablation shows LoRA slightly outperforming full fine-tuning, with 0.9033 versus 0.8989 on validation (Gao et al., 17 Jul 2025).

The test-set results are the paper’s principal performance claims.

Configuration	Final score
m0: mPLUG-Owl2-7B, 1024, Full	0.9098
m1: mPLUG-Owl2-7B, 1024, LoRA	0.9108
m3: mPLUG-Owl2-7B, KonIQ pretrain, 1024, LoRA	0.9112
Q0: Qwen2.5-VL-7B, single model	0.9054
Q1: Qwen2.5-VL-7B, 5-fold ensemble	0.9235
m0 + m1 + m3 + Q0 + Q1 average	0.9288

The best reported system is the final multi-model ensemble, which achieves Final Score = 0.9288, with Sharpness = 0.9275, Color Fidelity = 0.9198, and Overall = 0.9339. The progression of results indicates that the largest gains arise from four sources: moving from Q-Align to DeQA-Score-style soft distribution learning, increasing effective input resolution to 1024, choosing pseudo variance over interpolation, and applying model ensemble across backbones and training variants. Prompt ensemble, by contrast, adds almost no improvement (Gao et al., 17 Jul 2025).

The paper also reports that pretraining on KonIQ improves DIQA-5000 performance, which suggests some transferability of quality-related representations from natural-image IQA to document images. This suggests cross-domain quality knowledge can be useful, even though DIQA remains a document-specific problem (Gao et al., 17 Jul 2025).

6. Relation to the DIQA literature, scope, and limitations

Within the emerging DIQA literature, DeQA-Doc occupies a distinct position. It is an MLLM-based score-distribution regressor for document quality, whereas later work such as “DocIQ: A Benchmark Dataset and Feature Fusion Network for Document Image Quality Assessment” emphasizes a specialized no-reference DIQA model with layout fusion, feature fusion, and multi-rater prediction on DIQA-5000 (Ma et al., 21 Sep 2025). A plausible implication is that these works represent complementary design points: DeQA-Doc pushes language-aligned MLLM scoring, while DocIQ pushes document-structure-aware DIQA architecture.

Another relevant comparison is “Q-Doc: Benchmarking Document Image Quality Assessment Capabilities in Multi-modal LLMs”, which evaluates MLLMs on DIQA at coarse, middle, and fine granularity levels and reports inconsistent scoring, distortion misidentification, and severity misjudgment, while also showing that Chain-of-Thought prompting can improve performance (Huang et al., 14 Nov 2025). This context is important because it situates DeQA-Doc not simply as another DIQA model, but as part of a broader attempt to determine whether MLLMs can become reliable document-quality assessors rather than only general multimodal reasoners.

The paper also has clear limitations. It does not report a full cross-dataset generalization benchmark over multiple document-quality datasets. Its best performance depends on ensembles, which increase computational cost. The soft labels remain approximations because DIQA-5000 lacks variance annotations. The paper does not present a dedicated failure-case section, and it notes that prompt ensemble is not effective and that $i$5 can hurt performance. More broadly, the work is about document image quality assessment, not document-grounded question answering. This distinction matters because the phrase “DeQA-Doc” also appears descriptively in work on document-grounded QA over specialized technical documentation, particularly EDA tool manuals, where it denotes retrieval-backed question answering rather than image quality assessment (Pu et al., 2024). In the named framework discussed here, however, DeQA-Doc specifically denotes an adaptation of DeQA-Score for DIQA (Gao et al., 17 Jul 2025).

Taken together, these results position DeQA-Doc as a document-specific extension of MLLM-based image quality scoring whose key innovations are the soft-label redesign without variance annotations, resolution-flexible visual encoding for high-resolution document images, and ensemble-based robustness. Its empirical record indicates that score-distribution-based MLLM regression can transfer effectively from natural-image IQA to DIQA when adapted carefully for the structural and annotation peculiarities of document images (Gao et al., 17 Jul 2025).