AI Receipt Recognition: Data Extraction with 94% Accuracy

Accountants spend up to 30% of their working time on manual entry of data from cash register receipts and payment slips. Meanwhile, 3–5% of records contain errors: missing items, incorrect amounts, lost TINs. Receipts are the most complex document type for OCR: thermal paper fades over time, print quality is low, formats vary by retailer, and receipt structure is non-linear (items, discounts, totals can appear in any order). On the SROIE dataset, LayoutLMv3 achieves F1 0.974 — but that's on clean scanned documents. On real-quality mobile photos, it drops to 0.87–0.91. We have developed a pipeline that raises F1 to 0.95+ on real data through aggressive preprocessing and hybrid parsing (OCR + regex + LLM verification). Our experience includes 40+ projects on extracting data from receipts, invoices, and bills.

Why Preprocessing Is the Main Source of Errors

Mobile photo of a receipt suffers from: perspective distortion, blur, finger shadows, overexposure. Without correction, PaddleOCR makes errors on 15–25% of characters. After our pipeline, CER drops by a factor of 2–3: from 4–5% to 1.5% on good photos, from 18% to 6% on poor ones.

import cv2
import numpy as np
from PIL import Image

def preprocess_receipt_photo(
    image: np.ndarray,
    target_width: int = 768   # ширина нормализованного чека
) -> np.ndarray:
    """
    Шаги: шумоподавление → выравнивание яркости → deskew → binarization
    """
    # 1. Шумоподавление
    denoised = cv2.fastNlMeansDenoisingColored(
        image, h=10, hColor=10,
        templateWindowSize=7, searchWindowSize=21
    )

    # 2. CLAHE для выравнивания освещённости
    lab = cv2.cvtColor(denoised, cv2.COLOR_BGR2LAB)
    clahe = cv2.createCLAHE(clipLimit=3.0, tileGridSize=(8, 8))
    lab[:, :, 0] = clahe.apply(lab[:, :, 0])
    enhanced = cv2.cvtColor(lab, cv2.COLOR_LAB2BGR)

    # 3. Автоматическое определение контура чека и deskew
    gray = cv2.cvtColor(enhanced, cv2.COLOR_BGR2GRAY)
    edges = cv2.Canny(gray, 50, 150)
    contours, _ = cv2.findContours(
        edges, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE
    )

    if contours:
        # Находим самый большой контур — предположительно чек
        largest = max(contours, key=cv2.contourArea)
        hull = cv2.convexHull(largest)

        if cv2.contourArea(hull) > 0.1 * image.shape[0] * image.shape[1]:
            rect = cv2.minAreaRect(hull)
            angle = rect[2]
            if abs(angle) > 5:
                M = cv2.getRotationMatrix2D(
                    (image.shape[1]//2, image.shape[0]//2), angle, 1
                )
                enhanced = cv2.warpAffine(
                    enhanced, M, (image.shape[1], image.shape[0])
                )

    # 4. Адаптивная бинаризация для термопечати
    gray = cv2.cvtColor(enhanced, cv2.COLOR_BGR2GRAY)
    binary = cv2.adaptiveThreshold(
        gray, 255,
        cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
        cv2.THRESH_BINARY,
        blockSize=11, C=2
    )

    # 5. Нормализация ширины
    h, w = binary.shape
    scale = target_width / w
    resized = cv2.resize(
        binary, (target_width, int(h * scale)),
        interpolation=cv2.INTER_LANCZOS4
    )

    return resized

How to Parse the Receipt Structure

After OCR, we need to parse the structure: item lines, discounts, totals, cashier, TIN.

import re
from dataclasses import dataclass, field
from typing import Optional

@dataclass
class ReceiptLineItem:
    name: str
    quantity: float
    unit_price: float
    total_price: float
    discount: float = 0.0

@dataclass
class ParsedReceipt:
    store_name: Optional[str]
    inn: Optional[str]          # ИНН продавца
    datetime_str: Optional[str]
    items: list[ReceiptLineItem] = field(default_factory=list)
    subtotal: Optional[float] = None
    tax_amount: Optional[float] = None
    total: Optional[float] = None
    payment_method: Optional[str] = None
    fiscal_sign: Optional[str] = None

class ReceiptParser:
    # Паттерны для российских чеков (ФФД 1.05/1.1)
    PATTERNS = {
        'inn': r'ИНН\s*[:№]?\s*(\d{10,12})',
        'total': r'(?:ИТОГО|ИТОГО К ОПЛАТЕ|ИТОГ)\s*[:=]?\s*([\d\s,\.]+)',
        'tax': r'(?:НДС|В т\.ч\. НДС)\s+\d+%\s*[:=]?\s*([\d\s,\.]+)',
        'fiscal_sign': r'ФП\s*[::]?\s*(\d+)',
        'line_item': r'^(.+?)\s+([\d,\.]+)\s*[xх×]\s*([\d,\.]+)\s*=?\s*([\d,\.]+)',
        'datetime': r'(\d{2}[\.\/]\d{2}[\.\/]\d{4})\s+(\d{2}:\d{2}(?::\d{2})?)',
    }

    def parse(self, ocr_text: str) -> ParsedReceipt:
        lines = ocr_text.split('\n')
        receipt = ParsedReceipt(
            store_name=lines[0].strip() if lines else None,
            inn=self._extract(ocr_text, 'inn'),
            datetime_str=self._extract_datetime(ocr_text),
            total=self._parse_amount(self._extract(ocr_text, 'total')),
            tax_amount=self._parse_amount(self._extract(ocr_text, 'tax')),
            fiscal_sign=self._extract(ocr_text, 'fiscal_sign')
        )

        for line in lines:
            item = self._parse_line_item(line)
            if item:
                receipt.items.append(item)

        return receipt

    def _extract(self, text: str, key: str) -> Optional[str]:
        m = re.search(self.PATTERNS[key], text, re.IGNORECASE)
        return m.group(1).strip() if m else None

    def _extract_datetime(self, text: str) -> Optional[str]:
        m = re.search(self.PATTERNS['datetime'], text)
        if m:
            return f'{m.group(1)} {m.group(2)}'
        return None

    def _parse_amount(self, text: Optional[str]) -> Optional[float]:
        if not text:
            return None
        cleaned = re.sub(r'\s', '', text).replace(',', '.')
        try:
            return float(cleaned)
        except ValueError:
            return None

    def _parse_line_item(self, line: str) -> Optional[ReceiptLineItem]:
        m = re.match(self.PATTERNS['line_item'], line.strip())
        if not m:
            return None
        try:
            return ReceiptLineItem(
                name=m.group(1).strip(),
                quantity=float(m.group(2).replace(',', '.')),
                unit_price=float(m.group(3).replace(',', '.')),
                total_price=float(m.group(4).replace(',', '.'))
            )
        except (ValueError, AttributeError):
            return None

Comparison of Field Extraction Approaches

The hybrid approach provides optimal balance: accuracy close to LLM at 10x lower cost.

Comparison of OCR Engines

PaddleOCR v4 is 4.5x more accurate than Tesseract on poor photos and is completely free.

How We Do It: Stack and Process

Stack: PaddleOCR v4 for text detection, custom LayoutLMv3 model for field extraction, PyTorch + ONNX Runtime for CPU inference. For complex cases — GPT-4o as fallback. Position embedding vectorization via Hugging Face.

Process:

• Analysis of client's receipt sample (minimum 200 samples).
• Field annotation: date, total, TIN, items.
• Train/fine-tune model on annotated data.
• Integration via REST API or Python library.
• Pilot on real data — evaluate accuracy (F1) and speed (p99 latency).
• Deliver documentation, operator training, 1-month support.

Timeline: from 2 weeks for a single store chain, from 6 weeks for a universal solution.

What's Included

• Pre-trained model on your data
• REST API with documentation (Swagger)
• Employee training (2 hours)
• 1-month support
• Accuracy guarantee >95% on target formats

Why Choose Us

10+ years of experience in computer vision, 40+ OCR deployments in finance, logistics, retail. Certified partners of OpenAI and Microsoft. We guarantee F1 on your receipts will be no lower than 0.93, otherwise we refine for free.

Get in Touch

Contact us for a pilot project estimate. Get a consultation on your task — we will suggest the optimal stack and timeline.

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.