Developing an Image Augmentation Pipeline for Training a CV Model

We encountered a typical task: a dataset of 800 images became a dataset of 24,000 — sounds nice, but if augmentations are wrong, the model simply overfits to the artifacts of the transformations themselves. Augmentations should mimic real production data changes, not just increase the number of pictures. Our practice confirms: a properly configured pipeline boosts accuracy by 5–12% without new labeling and saves budget. For example, in a metal defect detection project we initially applied photometric augmentations, but the model started confusing scratches with glare. After switching to geometric and adding synthetic via Stable Diffusion, recall improved from 0.61 to 0.78 with precision above 0.80. Labeling savings ranged from $2,000 to $10,000. On average, clients save from $5,000 to $15,000 per project.

How to choose an augmentation policy for a specific domain?

Configuring augmentations depends on the domain and task. For outdoor, strong illumination and weather changes are needed. For medical, only geometric and cautious photometric. Below is a basic policy on Albumentations that we adapt for each project.

import albumentations as A
from albumentations.pytorch import ToTensorV2

def build_augmentation_pipeline(
    task: str = 'detection',    # 'classification' | 'detection' | 'segmentation'
    domain: str = 'outdoor'     # 'outdoor' | 'indoor' | 'medical' | 'satellite'
) -> A.Compose:
    """
    Augmentation policy depends on domain.
    Outdoor — aggressive illumination and weather changes.
    Medical — only geometric, color changes contraindicated.
    """
    geometric = [
        A.HorizontalFlip(p=0.5),
        A.ShiftScaleRotate(
            shift_limit=0.05 if task == 'detection' else 0.1,
            scale_limit=0.1,
            rotate_limit=15,
            border_mode=0,     # constant padding
            p=0.5
        ),
    ]

    photometric_outdoor = [
        A.RandomBrightnessContrast(
            brightness_limit=0.3, contrast_limit=0.3, p=0.6
        ),
        A.HueSaturationValue(
            hue_shift_limit=15, sat_shift_limit=40, val_shift_limit=30,
            p=0.4
        ),
        A.RandomFog(fog_coef_lower=0.1, fog_coef_upper=0.35, p=0.15),
        A.RandomRain(
            slant_lower=-10, slant_upper=10,
            drop_length=15, drop_width=1,
            drop_color=(200, 200, 200), blur_value=2,
            brightness_coefficient=0.85, p=0.1
        ),
        A.RandomSunFlare(
            flare_roi=(0, 0, 1, 0.5), angle_lower=0,
            num_flare_circles_lower=3, num_flare_circles_upper=6,
            src_radius=200, p=0.08
        ),
    ]

    photometric_medical = [
        A.RandomGamma(gamma_limit=(80, 120), p=0.4),
        A.CLAHE(clip_limit=3.0, tile_grid_size=(8, 8), p=0.3),
    ]

    photometric = (
        photometric_outdoor if domain == 'outdoor'
        else photometric_medical
    )

    noise = [
        A.GaussNoise(var_limit=(10, 50), p=0.3),
        A.ImageCompression(quality_lower=75, quality_upper=100, p=0.2),
        A.Blur(blur_limit=3, p=0.1),
    ]

    bbox_params = (
        A.BboxParams(
            format='yolo',
            label_fields=['class_labels'],
            min_visibility=0.3
        )
        if task == 'detection' else None
    )

    transforms = geometric + photometric + noise + [
        A.Normalize(
            mean=(0.485, 0.456, 0.406),
            std=(0.229, 0.224, 0.225)
        ),
        ToTensorV2()
    ]

    return A.Compose(transforms, bbox_params=bbox_params)

When to use MixUp and CutMix?

Simple augmentations do not eliminate overfitting on small datasets. MixUp and CutMix create new examples by mixing two images — this is good regularization. MixUp, proposed in the paper Zhang et al., 2018 MixUp: Beyond Empirical Risk Minimization, is better for tasks with global features, while CutMix is for local patterns. It is important to control the lambda coefficient: for alpha=0.4, the mixing is moderate, and the model generalizes better.

import torch
import numpy as np

def mixup_batch(
    images: torch.Tensor,     # (B, C, H, W)
    labels: torch.Tensor,     # (B,) for classification
    alpha: float = 0.4
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, float]:
    lam = np.random.beta(alpha, alpha)
    batch_size = images.size(0)
    perm = torch.randperm(batch_size)
    mixed_images = lam * images + (1 - lam) * images[perm]
    labels_a = labels
    labels_b = labels[perm]
    return mixed_images, labels_a, labels_b, lam

def cutmix_batch(
    images: torch.Tensor,
    labels: torch.Tensor,
    alpha: float = 1.0
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, float]:
    lam = np.random.beta(alpha, alpha)
    batch_size, _, H, W = images.shape
    perm = torch.randperm(batch_size)
    cut_ratio = np.sqrt(1 - lam)
    cut_h = int(H * cut_ratio)
    cut_w = int(W * cut_ratio)
    cx = np.random.randint(W)
    cy = np.random.randint(H)
    x1 = max(cx - cut_w // 2, 0)
    y1 = max(cy - cut_h // 2, 0)
    x2 = min(cx + cut_w // 2, W)
    y2 = min(cy + cut_h // 2, H)
    mixed = images.clone()
    mixed[:, :, y1:y2, x1:x2] = images[perm, :, y1:y2, x1:x2]
    lam = 1 - (y2-y1) * (x2-x1) / (H * W)
    return mixed, labels, labels[perm], lam

def mixup_criterion(criterion, pred, y_a, y_b, lam):
    return lam * criterion(pred, y_a) + (1 - lam) * criterion(pred, y_b)

Synthetic data with Stable Diffusion

Note: when real data has fewer than 100 examples per class, synthetic via inpainting or ControlNet gives an AP gain of 8–15%. We use Stable Diffusion Inpainting to generate defects on empty backgrounds. We also use ControlNet for contour generation if precise object insertion is needed.

from diffusers import StableDiffusionInpaintPipeline
import torch
from PIL import Image

pipe = StableDiffusionInpaintPipeline.from_pretrained(
    'runwayml/stable-diffusion-inpainting',
    torch_dtype=torch.float16
).to('cuda')

def generate_synthetic_defect(
    background_image: Image.Image,
    mask_image: Image.Image,
    defect_type: str = 'surface crack'
) -> Image.Image:
    prompt = (
        f'industrial {defect_type}, high resolution, '
        f'realistic texture, macro photography'
    )
    result = pipe(
        prompt=prompt,
        image=background_image,
        mask_image=mask_image,
        num_inference_steps=30,
        guidance_scale=7.5,
        strength=0.85
    ).images[0]
    return result

In practice: 80 real crack examples + 400 synthetic ones raised recall from 0.61 to 0.78 while maintaining precision above 0.80. Labeling savings are significant, often reaching $10,000.

Comparison of augmentation strategies

Typical mistakes when configuring augmentations

• Applying photometric augmentations to medical images — can destroy diagnostic features.
• Using overly aggressive geometric transformations for detection — bboxes may fall outside the image.
• Lack of validation set checks — augmentations can change data distribution.
• Skipping normalization after augmentations — breaks pretrained weights.
• Incorrect normalization after augmentations — a common cause of metric drops on validation.

What is included in our augmentation pipeline?

We provide:

• Dataset and domain analysis, identification of key variations.
• Strategy selection: geometry, photometry, MixUp/CutMix, synthetic data.
• Pipeline implementation on Albumentations, considering task (bbox, masks).
• Integration into training loop (PyTorch/TensorFlow).
• Experiments: A/B testing combinations, metric logging in Weights & Biases.
• Synthetic data generation (optional) via Stable Diffusion.
• Model deployment with final pipeline (inference-only transforms).
• Documentation and team training.

Process overview

1. Dataset and domain analysis: study data variability, typical production distortions.
2. Strategy selection: geometry, photometry, MixUp/CutMix, or synthetic.
3. Pipeline configuration on Albumentations considering task (bbox, masks).
4. Integration into training loop (PyTorch/TensorFlow).
5. Experiments: test different combinations, log metrics.
6. Synthetic data generation (optional) via Stable Diffusion.
7. Model deployment with final pipeline (inference-only transforms).

Timeline and pricing

Pricing is calculated individually. We guarantee a transparent process and metric tracking at each stage. Contact us to evaluate your project — we'll select the optimal end-to-end augmentation strategy. Order an audit of your current pipeline — get specific recommendations for improvement. 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.