Building HTR Systems for Handwriting: TrOCR and PaddleOCR

Handwritten text is significantly more complex than machine-printed: infinite variety of handwriting styles, ligatures (connected writing), blurry boundaries between characters, variability in pressure and angle. We encounter this on every project. Clients often come with a task: to recognize thousands of hand-filled forms per month with an error rate no higher than 5%. Standard OCR systems are powerless here. Over 5 years, we have implemented more than 30 projects for recognizing handwritten text for archival, medical, and corporate clients. Our experience confirms: proper preprocessing and choice of architecture solve 80% of the success. In one project for a medical center, we replaced manual entry of 5000 charts per day with automatic recognition — this reduced document processing costs by more than 60%, saving approximately $20,000 annually. Our HTR system achieved CER below 5% on these medical chart recognition tasks.

Choosing a Model for Russian Handwritten Text

Choosing an architecture depends on the language and data volume. TrOCR (Microsoft) is a transformer encoder-decoder for OCR. Encoder: ViT image processing, Decoder: autoregressive text generation. State-of-the-art on IAM handwriting dataset: CER 2.89% (large model). However, TrOCR is trained mainly on English, so for Cyrillic OCR it is better to use PaddleOCR with its SVTR_LCNet architecture, which leverages spatial transformer networks and attention mechanisms for robust recognition. PaddleOCR outperforms TrOCR on Cyrillic by a factor of 2 in CER, making it a better choice for Russian handwritten text.

from transformers import TrOCRProcessor, VisionEncoderDecoderModel
from PIL import Image
import torch

class HandwritingRecognizer:
    def __init__(self, model_name: str = 'microsoft/trocr-large-handwritten'):
        self.processor = TrOCRProcessor.from_pretrained(model_name)
        self.model = VisionEncoderDecoderModel.from_pretrained(model_name)
        self.model.eval()
        self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
        self.model.to(self.device)

    @torch.no_grad()
    def recognize(self, image: Image.Image) -> str:
        """Recognition of a single line of text"""
        pixel_values = self.processor(
            images=image,
            return_tensors='pt'
        ).pixel_values.to(self.device)

        generated_ids = self.model.generate(
            pixel_values,
            max_new_tokens=128,
            num_beams=4
        )

        return self.processor.batch_decode(
            generated_ids,
            skip_special_tokens=True
        )[0]

PaddleOCR for handwritten Cyrillic text significantly outperforms TrOCR:

from paddleocr import PaddleOCR

ocr = PaddleOCR(
    use_angle_cls=True,
    lang='ru',
    rec_algorithm='SVTR_LCNet',
    rec_model_dir='./models/handwriting_rec'
)

Why Preprocessing Matters More Than Architecture

Handwritten text requires more aggressive preprocessing. Removing ruled background, binarization, cleaning artifacts — these steps critically affect the final CER. Below is a typical handwriting preprocessing pipeline.

import cv2
import numpy as np
from skimage import morphology

def preprocess_handwriting(image: np.ndarray) -> np.ndarray:
    gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)

    # Remove background lines (ruled paper)
    horizontal_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (40, 1))
    horizontal_lines = cv2.morphologyEx(gray, cv2.MORPH_OPEN, horizontal_kernel)
    gray = cv2.subtract(gray, horizontal_lines)

    # Otsu binarization
    _, binary = cv2.threshold(gray, 0, 255,
                               cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)

    # Remove small artifacts
    cleaned = morphology.remove_small_objects(
        binary.astype(bool), min_size=50
    ).astype(np.uint8) * 255

    return cleaned

How We Perform Model Fine-Tuning

To adapt for corporate data, we use the following pipeline for neural network fine-tuning:

1. Data annotation. We collect 500–2000 lines of handwritten text, transcribe them in Label Studio. Each line is a separate file with a text label. Label Studio annotation ensures high-quality ground truth.
2. Augmentation. We apply random shift, rotation up to 5°, scaling 0.9–1.1, adding noise, and elastic distortion — this increases resilience to handwriting variations via self-supervised pretext tasks.
3. Fine-tuning. For PaddleOCR we use RecModel with SVTR_LCNet, batch size 32, learning rate 1e-4, 100 epochs, employing CTC loss for alignment. We monitor CER metric on validation.
4. Validation. We test on 10% of data (not used in training). If CER is above 10%, we add data or change hyperparameters.
5. Export. The model is converted to ONNX or saved in PaddleOCR format for inference with beam search decoding.

Document Line Segmentation for Multi-Line Documents

Before recognition, a multi-line document must be split into lines. Our document line segmentation uses horizontal projection:

def segment_lines(binary_image: np.ndarray) -> list[np.ndarray]:
    """Horizontal projection for line segmentation"""
    horizontal_projection = binary_image.sum(axis=1)

    threshold = horizontal_projection.max() * 0.05
    in_line = horizontal_projection > threshold

    lines = []
    start = None
    for i, active in enumerate(in_line):
        if active and start is None:
            start = max(0, i - 5)
        elif not active and start is not None:
            end = min(len(in_line), i + 5)
            line_img = binary_image[start:end, :]
            if end - start > 10:
                lines.append(line_img)
            start = None

    return lines

Fine-Tuning on Corporate Handwritten Data

For specific handwriting (medical records of a particular hospital, enterprise forms), fine-tuning is required. Without it, CER can reach 20-30%, which is unacceptable for document flow. We guarantee that after fine-tuning on 500–2000 lines, accuracy will increase to 90-95%. TrOCR fine-tuning typically requires a few hundred annotated images, while PaddleOCR fine-tuning for handwriting can leverage pre-trained checkpoints.

• Annotation of 500–2000 lines via Label Studio or CVAT.
• Fine-tuning TrOCR or PaddleOCR rec_model.
• CER decreases from 15–25% to 5–10% on domain-specific data.

Understanding CER (Character Error Rate)

CER (Character Error Rate) is the proportion of errors at the character level. We use the CER metric to evaluate accuracy. For business processes, even 5% can mean hundreds of incorrectly recognized digits in reports. In one project for a medical center, we reduced CER from 18% to 4% by applying a combination of adaptive binarization and PaddleOCR fine-tuning. The result — automatic processing of 5000 charts per day instead of manual entry, and reduction of document processing costs by more than 60%. Reference CER values on public datasets (IAM Handwriting Database) are ~2.89%, but on real data fine-tuning is required.

Model Update Frequency

If operator handwriting changes or new fields are added, the model should be retrained every six months. We provide an incremental learning pipeline that allows updating weights in a few hours without full retraining. When the document template changes significantly (e.g., switching to a new form), adding 200–500 new annotated lines and running fine-tuning is sufficient.

Work Deliverables

• Requirements analysis and test run on 50 pages.
• Architecture selection (TrOCR / PaddleOCR / combined).
• Development of preprocessing and segmentation pipeline.
• Model fine-tuning (if needed) with validation.
• Integration via REST API or gRPC.
• Documentation (API reference, deployment guide).
• Access to model repository and training scripts.
• Operator training session (up to 4 hours).
• 3 months support after deployment.

Timeline and Cost

Cost is calculated individually — depends on data volume, required accuracy, and integration complexity. Typical project cost: $5,000–$15,000 for a complete HTR system. In one deployment, we cut annual document processing costs by $20,000. Project evaluation is free. Contact us for a test run on your samples. Our engineers will analyze your handwritten documents and propose the optimal solution. Get a consultation today.

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.