PaddleOCR-VL Layer-12 Feature Extractor (ONNX)
Truncated PaddleOCR-VL vision encoder — 1152D intermediate features for document image quality assessment.
Layer 12 of the PaddleOCR-VL vision encoder captures mid-level visual patterns (texture, edge structure, compression artifacts, noise patterns) that are highly predictive of OCR readability. This ONNX model extracts those features with zero PyTorch dependency — just onnxruntime + numpy + Pillow.
Why Layer 12?
From a sensitivity analysis across all 27 encoder layers:
| Degradation | Best Layer | Spearman ρ | Notes |
|---|---|---|---|
| Gaussian Noise | layer_00 | 0.96 | Shallow layers detect fine noise |
| JPEG Compression | layer_08 | 0.86 | Mid layers sense blocking artifacts |
| Gaussian Blur | layer_08 | 0.66 | Moderate response |
| Downscale | pooled | 0.65 | Resolution loss detectable |
Layer 12 was chosen because:
- Highest intra/inter-class distance ratio (0.29) — features cluster by quality level, not degradation type
- Cross-resolution consistency >0.99 — same image at 224² vs 728² produces nearly identical features
- Good balance: deeper than noise-sensitive layers, shallower than semantic layers
Domain & Positioning: 这个模型做什么、不做什么
一句话定位
这是一个"OCR 场景的文档图片质量特征提取器"，不是通用图片质量打分器，也不是 OCR 识别器。
它的核心假设是：如果一张图在 PaddleOCR-VL 的中间层特征空间里和清晰参考图很接近，那这张图大概率能被 OCR 正确识别。
使用场景
| 场景 | 是否适合 | 说明 |
|---|---|---|
| 📄 文档上传质量校验 | ✅ 核心场景 | "这张扫描件够清晰吗？需要让用户重拍吗？" |
| 🏦 票据/证件 OCR 前置过滤 | ✅ 推荐 | 在送入 OCR 引擎前筛掉模糊/噪声过大的图片 |
| 📸 拍照文档质量评分 | ✅ 适合 | 模糊、噪声、JPEG 压缩 → 有强信号 |
| 🌐 通用图片美学评分 | ❌ 不适合 | 模型不理解"构图"、"色彩搭配"、"主体突出" |
| 🌃 自然场景图片质量 | ⚠️ 有限 | 对噪声/模糊有响应，但对过曝/偏色无感知 |
| 🔤 OCR 文字识别 | ❌ 做不了 | 这是特征提取器，不是 OCR 引擎。请用完整版 PaddleOCR-VL |
| 🎯 文字语义理解 | ❌ 做不了 | 深层语义信息在 layer 20+，layer 12 以纹理/结构信息为主 |
与同类模型的关键区别
1. vs 完整 PaddleOCR-VL（端到端 OCR 模型）
| 本模型 (Layer-12 ONNX) | 完整 PaddleOCR-VL | |
|---|---|---|
| 任务 | 图片质量特征提取 | 端到端 OCR（检测+识别） |
| 输出 | 1152D 特征向量 | 文字检测框 + 文本内容 |
| 推理速度 | ~100ms CPU | ~2-5s GPU |
| 依赖 | onnxruntime only | PyTorch + transformers |
| 部署 | 任意 ONNX 运行时（含移动端/边缘） | 需要 GPU 服务器 |
| 理解内容 | 不懂文字内容，只看纹理质量 | 完整理解文档内容 |
选择指南：如果你需要的是"这张图能读吗"的答案 → 用本模型提取特征 + 轻量回归器。如果你需要"图里写了什么" → 直接用完整 PaddleOCR-VL。
2. vs DINOv3（通用视觉特征提取器）
| 本模型 (Layer-12) | DINOv3 (ViT-S/16) | |
|---|---|---|
| 训练数据 | PaddleOCR-VL 图文数据（文档为主） | 大规模自然图片（ImageNet/LAION 级别） |
| 特征偏好 | 纹理、边缘、压缩伪影（文档质量相关） | 语义类别、物体形状、全局布局 |
| 文档噪声敏感度 | ⭐⭐⭐⭐⭐ 极高（ρ=0.96 for noise） | ⭐⭐ 中等（对噪声有响应但不如专用模型） |
| 亮度/对比度感知 | ⭐ 弱（盲区） | ⭐⭐ 中等 |
| 特征维度 | 1152D（单层） | 384D (ViT-S) ~ 1536D (ViT-L) |
关键差异：DINOv3 是"通才"，擅长判断"这是什么物体"；本模型是"文档专才"，擅长判断"这文档还能读吗"。在文档质量任务上，本模型的特征 + 轻量回归器（84K 参数）达到 ρ=0.74，优于直接对 DINOv3 特征做 XGBoost。
实测对比（合成文档质量数据集，12 种退化）：
| 特征源 + 回归器 | Spearman ρ | R² |
|---|---|---|
| PaddleOCR Layer-12 + Two-Tower MLP | 0.74 | 0.44 |
| PaddleOCR Layer-12 + XGBoost | 0.60 | 0.35 |
| CV 传统特征 only + XGBoost | 0.66 | 0.33 |
| DINOv3 ViT-S + XGBoost | ~0.50 | ~0.25 |
3. vs Qwen3-VL-Embedding / CLIP Embedding（通用 VLM 特征）
| 本模型 (Layer-12) | Qwen3-VL-Embedding / CLIP | |
|---|---|---|
| 特征来源 | 视觉编码器中间层（layer 12/27） | 视觉+文本对齐后的池化输出 |
| 语义层次 | 中层纹理/结构（edges, textures, artifacts） | 高层语义（"这是一张包含表格的文档"） |
| 跨模态 | 纯视觉，无文本对齐 | 视觉-文本对齐（CLIP 训练目标） |
| 质量退化感知 | ⭐⭐⭐⭐⭐ 天然敏感 | ⭐⭐ "模糊的文档"和"清晰的文档"在语义空间可能很近 |
| 相似度计算 | 余弦距离 ≈ 退化程度 | 余弦相似度 ≈ 语义相似度 |
核心差异：CLIP/Qwen3-VL 的 embedding 回答"这两张图内容像不像"；本模型的特征回答"这张图和清晰版差多远"。对于文档质量任务，你不想要语义相似度——模糊的身份证和清晰的身份证在语义上是同一张证，但在质量维度上区别巨大。
4. vs Q-ReAlign（通用 VLM 质量评分）
Q-ReAlign 是直接用 VLM 做质量评分的方案（将质量分数映射到离散 token 做软分类）。
| 本模型 (Layer-12) | Q-ReAlign | |
|---|---|---|
| 方法论 | 特征提取 + 下游回归器 | 端到端 VLM 质量打分 |
| 文档质量表现 | ρ=0.74 | ρ ≈ -0.60（对 blur/JPEG 评分反向） |
| 原因 | 特征空间天然反映退化程度 | 训练数据是自然照片美学评分，不适用于文档 |
| 优势 | 文档场景专用，轻量部署 | 通用场景，零样本即可用 |
选择指南：Q-ReAlign 适合自然照片（"这张风景照好看吗"），不适合文档（"这张扫描件能 OCR 吗"）。
5. vs 传统 CV 质量指标（Laplacian / FFT / Canny）
| 本模型 (Layer-12) | 传统 CV 指标 | |
|---|---|---|
| 噪声检测 ρ | 0.96 | ~0.30 (Laplacian 对噪声不敏感) |
| 模糊检测 ρ | 0.72 | ~0.85 (Laplacian 天然适合) |
| 压缩伪影 ρ | 0.86 | ~0.20 (无对应指标) |
| 亮度异常 | 0.05 (盲区) | ~0.80 (直方图直接捕获) |
| 可解释性 | 黑盒向量 | ✅ 每个指标有物理含义 |
| 部署成本 | ONNX Runtime ~100ms | 零成本，纯数学运算 <1ms |
结论：两者互补。传统 CV 覆盖亮度/对比度（模型的盲区），模型覆盖噪声/压缩（CV 的盲区）。最佳方案是双塔融合：本模型 1152D + CV 6D → 联合回归器。
模型选型速查表
| 你的需求 | 推荐方案 |
|---|---|
| 文档上传前质量检查 | 本模型 (feature extractor) + 轻量 Two-Tower 回归器 |
| 通用图片质量打分（非文档） | Q-ReAlign / CLIP-IQA / MUSIQ |
| 端到端 OCR 识别 | 完整 PaddleOCR-VL / PaddleOCR Server |
| 零依赖快速模糊检测 | 传统 CV (Laplacian variance) |
| 图片语义搜索/相似度 | CLIP / Qwen3-VL-Embedding / DINOv3 |
| 移动端文档质量 | 本模型 ONNX (支持 ONNX Runtime 全平台) |
Model Architecture
Input Image (any size)
│
▼
PaddleOCRVLImageProcessor (smart_resize + normalize + patchify)
│
▼
Patch Embedding + Position Encoding
│
▼
Transformer Encoder Layers 0 → 12 ← TRUNCATED HERE (original: 27 layers)
│
▼
Mean Pooling across patches
│
▼
1152-D Feature Vector
| Property | Value |
|---|---|
| Parameters | 237M (905 MB fp32 weights) |
| Input | pixel_values: (1, N, 3, 14, 14) float32, position_ids: (1, 1) int64 |
| Output | layer_12_pooled: (1, 1152) float32 |
| Dynamic axes | num_patches (varies with input resolution) |
| Opset | 18 |
| Files | model.onnx (851 KB graph) + model.onnx_data (902 MB weights) |
Installation
pip install onnxruntime numpy Pillow opencv-python
No PyTorch, no transformers, no CUDA required. CPU inference is ~50-200ms per image.
Quick Start
from inference.onnx_inference import Layer12ONNXExtractor
from PIL import Image, ImageFilter
# Load model
extractor = Layer12ONNXExtractor("model.onnx")
# Extract features
img = Image.open("document.jpg").convert("RGB")
features = extractor.extract(img) # → np.ndarray shape (1152,)
# Quality via distance from pristine reference
pristine = Image.open("pristine.jpg").convert("RGB")
blurred = img.filter(ImageFilter.GaussianBlur(radius=5))
quality = extractor.quality_score(blurred, reference=pristine)
# → 0.0 (heavily degraded) ~ 1.0 (pristine)
See inference/example.py for a complete walkthrough.
Feature Quality Benchmarks
Run with: python benchmark/run_benchmark.py
1. Degradation Sensitivity (Spearman ρ)
Correlation between feature distance and degradation severity across 12 degradation types × 7 levels × 5 image types:
| Degradation | Spearman ρ | Monotonicity | Sensitivity |
|---|---|---|---|
| Gaussian Noise | 0.96 | 0.98 | Excellent |
| JPEG Compression | 0.86 | 0.97 | Excellent |
| Gaussian Blur | 0.72 | 0.95 | Good |
| Motion Blur | 0.68 | 0.93 | Good |
| Median Blur | 0.65 | 0.92 | Good |
| Downscale | 0.65 | 0.90 | Good |
| ISO Noise | 0.60 | 0.88 | Fair |
| Salt & Pepper | 0.55 | 0.85 | Fair |
| Rotation | 0.30 | 0.60 | Weak |
| Brightness | 0.05 | 0.52 | Blind spot |
| Contrast | 0.05 | 0.50 | Blind spot |
Mean |ρ| = 0.55 across all degradations.
2. Paired Ranking Accuracy
Given a pristine reference and two degraded copies at different severity levels, can the model correctly rank which is worse?
Overall: 89% across 200 random pairs.
3. Cross-Resolution Consistency
Same content at different resolutions → same features? Cosine similarity between features extracted at 224² vs 728²:
Mean: 0.997 — near-perfect consistency thanks to PaddleOCR-VL's smart_resize.
Use Cases
- Document upload quality gate — reject blurry/noisy scans before OCR
- Image preprocessing quality monitor — detect when enhancement pipeline degrades
- Reference-based quality scoring — compare against a known-good template
- Feature backbone for quality regression — use 1152D features as input to a lightweight quality regressor (e.g., Two-Tower MLP, XGBoost)
Limitations
- Blind to brightness/contrast issues — these don't affect the feature space. Complement with traditional CV metrics (Laplacian variance, histogram stats).
- Geometric degradations (rotation) have weak signal — the model uses position encoding, but rotated text still looks like text to mid-layer features.
- Not a standalone quality scorer — this is a feature extractor. You need a downstream head (trained regressor, reference comparison, or CV complement) for final quality scores.
- Original PaddleOCR-VL base model — the vision encoder is from PaddleOCR-VL, which is optimized for Chinese + English document OCR.
Recommended Quality Pipeline
For a robust document quality system, use a Two-Tower fusion:
PaddleOCR-VL Layer 12 (this model) OpenCV traditional metrics
│ │
PCA → 128D 6D vector
Deep Tower MLP CV Tower MLP
│ │
└──────── concat (80D) ───────────────┘
│
Shared Head
│
quality_score [0, 1]
This approach achieves ρ = 0.74 on held-out data (vs 0.60 for XGBoost on raw features).
Open-Source Benchmarks for Evaluation
If you want to benchmark your quality model against published work:
Direct: OCR Quality Assessment
| Dataset | Size | Focus | Link |
|---|---|---|---|
| OCR-Quality | 1,000 images | Human-annotated OCR quality (4-level) | Aslan-mingye/OCR-Quality |
| DocPTBench | 1,300+ photos | Photographed document quality (geometric + photometric) | Topdu/DocPTBench |
General Image Quality Assessment (IQA)
| Dataset | Size | Focus | Link |
|---|---|---|---|
| LIVE | 779 images | 5 distortion types, DMOS labels | UT Austin LIVE |
| CSIQ | 866 images | 6 distortion types, DMOS labels | CSIQ |
| TID2013 | 3,000 images | 24 distortion types, MOS labels | TID2013 |
| KADID-10k | 10,125 images | 25 distortion types × 5 levels | KADID-10k |
| KonIQ-10k | 10,073 images | In-the-wild quality, MOS labels | KonIQ-10k |
OCR Recognition Benchmarks
| Dataset | Size | Focus | Link |
|---|---|---|---|
| GlotOCR Bench | 158 scripts | Multilingual OCR with degraded variants | HuggingFace (eval license) |
| Misraj-DocOCR | 400 images | Arabic document OCR (WER/CER) | Open-source |
| OCRGenBench | 1,060 samples | OCR generation quality | PeirongZhang/OCRGenBench |
Recommended starting point: OCR-Quality — it's the only dataset that directly measures OCR quality assessment, with human annotations and per-page MOS scores. Our benchmark script includes a loader for it.
Files
paddleocr-quality-onnx/
├── README.md # This file
├── .gitattributes # HF LFS config
├── requirements.txt # Python dependencies
├── model.onnx # ONNX model (~0.8 MB)
├── inference/
│ ├── __init__.py
│ ├── preprocessing.py # Standalone image preprocessing (no torch)
│ ├── onnx_inference.py # Layer12ONNXExtractor class
│ └── example.py # Complete usage example
└── benchmark/
├── run_benchmark.py # Degradation sensitivity, ranking, resolution
└── results/ # Benchmark output
Citation
If you use this model in your research:
@software{paddleocr-quality-onnx,
title = {PaddleOCR-VL Layer-12 Feature Extractor (ONNX)},
year = {2025},
note = {Truncated PaddleOCR-VL vision encoder for document image quality assessment},
url = {https://huggingface.co/[your-username]/paddleocr-quality-onnx},
}
The original PaddleOCR-VL model is from PaddlePaddle. This is a derived work using only the vision encoder (layers 0-12), exported to ONNX for lightweight deployment.
License
This model is derived from PaddleOCR-VL, which is released under Apache 2.0. This ONNX export and accompanying code are also Apache 2.0.