OCR pipeline for document photos

OCR pipeline for document photos

A photo of a document is not a scan. Typical problems: perspective distortion (text becomes trapezoidal), shadows from fingers and spine, uneven lighting, motion blur, reflections on glossy surfaces. A high-quality recognition system must correct all these distortions before sending to an OCR engine. Without preprocessing, even the best models (PaddleOCR, GPT-4o) show CER of 20–30% on such photos. For passports and IDs, CER <5% is required — achievable only with a comprehensive pipeline.

We developed a solution that automatically detects the document, aligns perspective, removes shadows and glare, and only then runs OCR. This approach yields consistently low error even on imperfect shots. Our pipeline processes a frame in ~200 ms on GPU and has been deployed in projects for banks and government organizations — over 5 years of experience, more than 20 successful integrations.

How we align the document in the photo?

The first task is to find the document in the frame and correct perspective. Typical situation: a user photographs a passport at an angle, and the text becomes unreadable for regular OCR. We use contour detection after preprocessing: gray conversion, blur, Canny edge detection, then search for a quadrilateral that occupies >20% of the frame. If such a contour is found, we apply perspective transformation (homography). This gives a front-facing view of the document.

import cv2
import numpy as np

class DocumentPhotoOCR:
    def __init__(self):
        self.ocr = PaddleOCR(use_angle_cls=True, lang='ru', use_gpu=True)

    def process(self, image_path: str) -> dict:
        image = cv2.imread(image_path)

        # 1. Detect document in frame
        doc_corners = self.detect_document(image)

        # 2. Perspective correction
        if doc_corners is not None:
            image = self.four_point_transform(image, doc_corners)

        # 3. Image enhancement
        image = self.enhance_document(image)

        # 4. OCR
        result = self.ocr.ocr(image, cls=True)

        return {
            'text': self._extract_text(result),
            'words': self._extract_words_with_positions(result),
            'corrected': doc_corners is not None
        }

    def detect_document(self, image: np.ndarray) -> np.ndarray | None:
        gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
        blur = cv2.GaussianBlur(gray, (5, 5), 0)
        edges = cv2.Canny(blur, 75, 200)
        dilated = cv2.dilate(edges, np.ones((3, 3)), iterations=1)
        contours, _ = cv2.findContours(dilated, cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)
        contours = sorted(contours, key=cv2.contourArea, reverse=True)[:5]
        for contour in contours:
            peri = cv2.arcLength(contour, True)
            approx = cv2.approxPolyDP(contour, 0.02 * peri, True)
            if len(approx) == 4:
                area_ratio = cv2.contourArea(approx) / (image.shape[0] * image.shape[1])
                if area_ratio > 0.2:
                    return approx.reshape(4, 2)
        return None

    def four_point_transform(self, image: np.ndarray, pts: np.ndarray) -> np.ndarray:
        rect = self._order_points(pts)
        tl, tr, br, bl = rect
        width = int(max(np.linalg.norm(br - bl), np.linalg.norm(tr - tl)))
        height = int(max(np.linalg.norm(tr - br), np.linalg.norm(tl - bl)))
        dst = np.array([
            [0, 0], [width - 1, 0],
            [width - 1, height - 1], [0, height - 1]
        ], dtype='float32')
        M = cv2.getPerspectiveTransform(rect.astype('float32'), dst)
        return cv2.warpPerspective(image, M, (width, height))

Result: even at tilt up to 30°, text becomes horizontal, which radically improves recognition quality.

What about shadows and glare?

The next problem is uneven lighting. Real photos often have shadows from fingers, binding, or glare on laminated cards. For shadows, we apply CLAHE (Contrast Limited Adaptive Histogram Equalization) in LAB color space. This adaptively equalizes brightness locally without creating artifacts. For glare, we use inpainting — detect overexposed pixels (value >250 in any RGB channel) and interpolate them from neighboring areas.

def enhance_document(self, image: np.ndarray) -> np.ndarray:
    lab = cv2.cvtColor(image, cv2.COLOR_BGR2LAB)
    l, a, b = cv2.split(lab)
    clahe = cv2.createCLAHE(clipLimit=4.0, tileGridSize=(16, 16))
    l_enhanced = clahe.apply(l)
    enhanced = cv2.merge([l_enhanced, a, b])
    enhanced = cv2.cvtColor(enhanced, cv2.COLOR_LAB2BGR)
    kernel = np.array([[0, -1, 0], [-1, 5, -1], [0, -1, 0]])
    sharpened = cv2.filter2D(enhanced, -1, kernel)
    return sharpened

We also add a light sharpen to compensate for blur from handheld shots. Our experience shows that the combination of CLAHE + sharpen reduces CER by 3–5% compared to raw images.

Full pipeline for document photo processing

Here is our production pipeline (Python code with OpenCV and PaddleOCR). The functions are combined into one class. We use PaddleOCR with Russian language and angle classification (use_angle_cls=True). On GPU, processing one frame takes ~200 ms. For mobile devices, a lightweight model can be substituted.

Comparison of OCR engines on real photos

PaddleOCR offers the best balance of accuracy and speed for Russian-language documents. Without our preprocessing pipeline, any OCR engine shows 5–10% worse CER. Using the combination of OpenCV and PaddleOCR, we achieve CER <5% for passports — 1.5–3 times better than without preprocessing. Our pipeline processes a frame 2x faster than standard solutions with similar quality.

Implementation process

Turnkey implementation includes:

1. Analysis: examining document types, shooting conditions, accuracy requirements, selecting reference images.
2. Design: stack selection (OpenCV, PyTorch/TensorFlow, OCR engine), pipeline architecture, container configuration.
3. Development: writing detection, correction, OCR modules; integration with backend via REST API.
4. Testing: on your dataset (minimum 500 images), measuring CER, latency p99, debugging edge cases.
5. Deployment: containerization (Docker), deployment on server or edge device, metric monitoring.

ocr = PaddleOCR(use_angle_cls=True, lang='ru', use_gpu=True, det_db_thresh=0.3, det_db_box_thresh=0.5)

What is included in the result

• Documented pipeline (code, configs, monitoring dashboards).
• Integration via REST API with request examples.
• Training webinar for your team.
• Guarantee: we set a target CER in the contract — your business gets predictable quality.

Estimated timelines

Pricing is calculated individually after analyzing your project. Contact us for an estimate — we'll discuss the stack and scope, prepare a commercial proposal in 1–2 days. Get an engineer consultation to ensure the solution is right for you.

We have been working with computer vision and OCR for over 5 years, delivering projects for banks, insurance companies, and government organizations. Your case could be next.

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.