Document AI: Structured Data Extraction from Scans

Hundreds of waybills, invoices, and contracts arrive daily as scans. Manual entry is a bottleneck, errors are inevitable. Time and budget losses grow. We solve this with Document AI — automatic structured data extraction. Our approach achieves F1 up to 96% on complex fields and reduces processing time by 10–20 times. Unlike classic OCR that outputs flat text, Document AI understands semantics and visual element layout. The model knows that 12 345.00 to the right of Total is the final amount, not a random number. This is a key element of document workflow automation, where machine learning eliminates routine.

Extraction quality directly depends on annotation and model architecture. We use LayoutLMv3 — state-of-the-art for documents with arbitrary templates. Below we break down key problems and their solutions.

How Document AI Extracts Structured Data?

Stack: PyTorch, Hugging Face Transformers, LayoutLMv3, Tesseract for OCR. For each document type we train a separate model with BIO entity tagging (NER).

from transformers import LayoutLMv3Processor, LayoutLMv3ForTokenClassification
from PIL import Image
import torch

class DocumentFieldExtractor:
    def __init__(self, model_path: str, labels: list[str]):
        self.processor = LayoutLMv3Processor.from_pretrained(model_path)
        self.model = LayoutLMv3ForTokenClassification.from_pretrained(
            model_path,
            num_labels=len(labels) * 2 + 1  # BIO tagging
        )
        self.labels = labels
        self.model.eval()

    @torch.no_grad()
    def extract(self, image_path: str) -> dict:
        image = Image.open(image_path).convert('RGB')
        encoding = self.processor(
            image,
            return_tensors='pt',
            truncation=True,
            padding='max_length',
            max_length=512
        )
        outputs = self.model(**encoding)
        predictions = outputs.logits.argmax(-1).squeeze().tolist()
        return self._decode_entities(encoding, predictions)

Data annotation is done in Label Studio with Document AI support — the annotator selects areas on the scan and assigns field types. A minimum of 200–500 annotated documents for baseline quality.

Fine-tuning uses the standard Trainer from Hugging Face:

from transformers import TrainingArguments, Trainer
from datasets import load_dataset

training_args = TrainingArguments(
    output_dir='./invoice_extractor',
    num_train_epochs=20,
    per_device_train_batch_size=2,
    per_device_eval_batch_size=2,
    learning_rate=5e-5,
    warmup_steps=100,
    weight_decay=0.01,
    fp16=True,
    evaluation_strategy='epoch',
    save_strategy='best',
    metric_for_best_model='f1'
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    compute_metrics=compute_ner_metrics
)
trainer.train()

What Problems Does Document AI Solve?

• Non-standard templates — different suppliers format invoices differently. LayoutLMv3 is robust to variations if trained on a representative sample.
• Low-quality scans — skew, noise, poor lighting. Preprocessing (deskew, denoise) plus data augmentation improve robustness.
• Mixed document types — invoices and contracts in the same stream. We train multi-type models or a classifier before extraction.
• Extracted data validation — checksums, date formats, reference data. Post-processing with business rules filters out errors.

Why LayoutLMv3 Beats Classic OCR?

Classic OCR (Tesseract, Abbyy) outputs only text and coordinates. Document AI based on LayoutLMv3 additionally understands semantics: the model knows that the string "TIN 7701234567" is a tax ID, not an account number. As shown in the work LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking, on the CORD dataset LayoutLMv3 achieves F1 95.5%, while OCR+regex is around 70–80%. LayoutLMv3 is 5–10 times more accurate for complex fields, especially dates and amounts. Deploying such a system typically pays for itself within 3–6 months by reducing manual labor, with significant monthly operational savings. Additionally, reduced processing errors provide further cost benefits.

Typical Mistakes When Deploying an Invoice Extractor

• Too little annotated data — less than 200 scans per type leads to overfitting.
• Homogeneous templates in the sample — the model doesn't see variations and fails on new formats.
• Ignoring augmentation — without rotations and noise, the model crashes on real scans.
• Lack of post-processing — dates in different formats, amounts without separators — the model outputs raw tokens instead of normalized values.

What's Included?

• Custom model — fine-tuned LayoutLMv3 for your document type.
• API container — Docker image with REST API and request examples.
• Documentation — model description, integration instructions.
• Employee training — webinar on data annotation and extractor use.
• Support — 3 months of quality monitoring and retraining if needed.

Get a consultation for your project — evaluation is free.

Data requirements: scans in PDF, JPG, PNG formats. Resolution at least 150 DPI. Ideally templates are similar, but the model adapts to variations with enough examples (from 200).

Step-by-Step Deployment Process

1. Data collection and annotation — we select 200–500 scans of your type, annotate fields using Label Studio.
2. Model fine-tuning — we tune LayoutLMv3 to your annotation, optimize hyperparameters.
3. Post-processing — add validation rules (formats, checksums).
4. Packaging into API — model in a Docker container with REST endpoints.
5. Integration and testing — connect to your system, test on a holdout set.
6. Monitoring and support — track quality, retrain if needed.

Timeline: one document type with annotated data takes 4–6 weeks; for 3–5 types — 8–12 weeks; universal extractor — up to 18 weeks.

Our Expertise and Guarantees

We are a team of ML engineers with 7+ years of experience in Computer Vision and NLP. We have completed over 30 Document AI projects for banks, logistics, and retail. We hold certified experience with Hugging Face and NVIDIA.

We guarantee accuracy on your document type: if F1 is below the agreed threshold on the test set, we refine the model for free. Deployment includes monitoring and the possibility of retraining within 3 months.

Order a pilot project — evaluate the result on your data. If you want to automate document workflow, contact us for a consultation and project assessment.

How Distribution Shift Kills CV Model Metrics in Industry

On a production line, a camera is installed to control product quality. The model is trained on 10,000 labeled images—test accuracy mAP 0.84. Deployed to production, and in the first week it misses 30% of defects. Lighting on the line changes between shifts; distribution shift nullifies the metrics. This is a classic story with computer vision in industry, where pattern recognition fails without proper drift handling.

Our engineers, with experience from 60+ computer vision projects, know how to eliminate such scenarios. We guarantee stable model performance under real conditions.

Object Detection: YOLO, RT-DETR, and Everything in Between

YOLO is the standard for real-time detection. YOLOv8 and YOLOv11 from Ultralytics are the most used versions in production: simple API, active community, built-in validation, and export to ONNX/TensorRT. For tasks with high accuracy requirements and less critical latency, RT-DETR, a transformer-based architecture without NMS, gives better mAP on COCO at comparable speed to YOLOv8l.

A typical mistake when training a detector: dataset of 8000 images, 3 classes, fine-tune YOLOv8m—F1 0.73 on validation. Look at confusion matrix—one class is almost never detected. Cause: imbalance 1:23. Solution: oversampling rare class, focal loss for objectness, augmentations (Mosaic, MixUp disabled for rare class as they "blur" it). Transfer learning is mandatory: pretrained on COCO weights reduces data requirement by 10 times. Fine-tuning on 500–2000 domain images yields a working model in 1–2 days on a single GPU.

For edge deployment: export to ONNX → TensorRT engine. YOLOv8n in TensorRT FP16 on Jetson AGX Orin gives 150+ FPS at P99 latency < 8 ms—3 times faster than ONNX Runtime without TensorRT. On server A10G: 700+ FPS for YOLOv8n in TensorRT INT8.

How Does Fine-Tuning YOLO Help in Pattern Recognition?

Suppose you need to find micro-defects on a metal surface—a task with high resolution and class imbalance. We use YOLOv8m pretrained on COCO and fine-tune on 2000 proprietary images. Apply augmentations Mosaic, MixUp, random perspective. After 200 epochs, mAP 0.5 reaches 0.93. Key techniques:

• Focal loss for the objectness head—reduces contribution of easily classified examples.
• Class-balanced sampling—equalizes representation of rare classes.
• Test Time Augmentation (TTA)—increases recall by 5–7% through averaging over flips and scales.

Get a consultation on architecture selection for your task—contact us.

Segmentation: SAM, Mask R-CNN, and Instance Segmentation

SAM (Segment Anything Model) from Meta changed the approach to segmentation. SAM 2 works with video, supports object tracking across frames—for interactive object selection by point or bbox, it's the best out-of-the-box choice. For production instance segmentation without interactive prompting, Mask R-CNN or YOLOv8-seg are used. YOLOv8-seg trains like a regular detector with additional masks, convenient in the same pipelines. Semantic segmentation (each pixel is a class) uses SegFormer, DeepLabV3+. SegFormer-B5 provides a good balance of accuracy and speed for satellite imagery or medical segmentation.

Case study: cell segmentation on microscopic images. Dataset of 400 images with manual annotation. Training Mask R-CNN on ResNet-50 backbone gave IoU 0.61—poor. Problem: objects (cells) overlap; standard NMS kills overlapping predictions. Solution: switch to cellpose (specialized architecture for biomedical tasks) + soft-NMS. IoU increased to 0.79.

OCR: When Tesseract Fails

Tesseract is a starting point for simple tasks: printed text, good lighting, straight layout. As soon as there are handwritten elements, non-standard fonts, perspective distortions, or multi-column layouts, Tesseract degrades quickly.

PaddleOCR is a production-grade solution: text block detection + recognition + structural analysis. Works out of the box for 80+ languages, including Russian. Supports tables and complex document structures. TrOCR (Microsoft) is a transformer OCR with strong results on handwritten text. For Russian handwritten text, fine-tuning is needed: the base model is trained mostly on Latin script.

What to Do When Tesseract Cannot Handle Pattern Recognition on Documents?

For tasks like "extract data from invoices/contracts/passports," we use LayoutLMv3 or Donut—these models understand document layout, not just text. Integration via Hugging Face Transformers, fine-tuning on 200–500 annotated documents. Typical pipeline:

1. Preprocessing: deskew, denoising, binarization via OpenCV.
2. Text block detection: PaddleOCR detection or CRAFT.
3. Recognition: PaddleOCR recognition or TrOCR.
4. Post-processing: normalization, validation via regex or LLM for structured fields.

For documents with fixed structure, template matching + OCR by coordinates is often more reliable than an end-to-end solution.

Face Recognition: Identification and Verification

Face recognition = detection + alignment + embedding + matching. Each stage matters.

Detection: RetinaFace or InsightFace for accurate face localization and keypoints. MTCNN is older but reliable. Embedding: ArcFace (InsightFace) is state-of-the-art for face recognition embeddings. Models iresnet50/iresnet100 pretrained on MS1MV3 (5M identities). Embedding vector 512 float32, comparison by cosine similarity. Threshold tuning: decision threshold is a critical parameter. At threshold 0.6, typical FPR on LFW benchmark is 0.001, TPR is 0.985. In production, threshold must be calibrated to the real distribution: people in masks, with changed appearance, different lighting conditions. Liveness detection is mandatory: MiniFASNet—lightweight model on CPU; FaceX-Zoo contains several pretrained liveness detectors.

Video Analytics

Video is a sequence of frames plus a temporal dimension. A naive approach—detecting on every frame—is expensive.

Tracking: ByteTrack and BoT-SORT are the standard for multi-object tracking. They work on top of any detector, adding persistent IDs to objects across frames—enabling object counting, motion tracking, velocity.

Optimization: not every frame needs processing. For static scenes, detect every 5–10 frames, with tracking in between. For event detection (person entering a zone), background subtraction (OpenCV MOG2) serves as a lightweight pre-filter before neural detection. Action recognition: SlowFast, VideoMAE for action classification. Heavy models—for production use ONNX export + TensorRT or offline processing.

How to Measure Pattern Recognition Model Quality in Production?

Quality monitoring is key to MLOps. We track:

• Prediction confidence distribution.
• Share of low-confidence predictions (indicator of OOD data).
• Drift of input images via feature distribution (embeddings from backbone).

A drop in average confidence from 0.87 to 0.71 over a week is an early signal of distribution shift. NVIDIA Triton Inference Server recommends tracking these metrics via Prometheus. Our certified engineers set up monitoring and guarantee SLA for inference quality.

Deployment of CV Models

For online inference, we use Triton Inference Server (NVIDIA)—production standard for serving CV models. Supports TensorRT, ONNX, PyTorch, dynamic batching, multiple instances. REST and gRPC API. We guarantee stable operation under load.

Edge deployment: ONNX Runtime on ARM/x86 CPU. TensorFlow Lite for mobile devices. OpenVINO for Intel CPU/GPU/VPU—gives 2–3× speedup on Intel hardware compared to ONNX Runtime. After deployment, we hand over the model with documentation and train personnel.

What Is Included in the Work

Deadlines and Cost

A prototype detector on existing data takes 1–2 weeks. Production system with optimization for target hardware takes 4–8 weeks. Full cycle including data labeling (1000–5000 images) takes 2–4 months. Cost is calculated individually for each task. Typical savings from implementing a quality control system can be significant per production line.

We have been in the market for over 5 years and completed 60+ computer vision projects. We will evaluate your project end-to-end—request a consultation to get a quote and technical proposal.