AI Super-Resolution — Image Upscaling

We constantly face the task: delivering maximum detail from a low-resolution source image. Bicubic interpolation gives 4x upscaling, but the image remains blurry, losing textures. AI super-resolution (Super-resolution) using Real-ESRGAN and GFPGAN solves this: it restores hair, text on signs, fabric structure. The difference is visible to the naked eye and in numbers: bicubic PSNR 28–30 dB, Real-ESRGAN 32–36 dB on photos. Modern models are trained on synthetic degradations, providing robustness to real noise and compression.

For commercial projects, the choice of model determines not only quality but also inference speed. Clients often come with old archives where resolution does not exceed 480p and want 4K for printing. We select a configuration that fits a reasonable budget: balancing detail and processing time.

For example, for an e-commerce client, we processed 50,000 product images: after upscaling, conversion increased by 15% thanks to better detail. The cost of integrating a ready-made solution is significantly lower than developing from scratch: on average, our clients save 60–80% of the budget.

How we implement upscaling for your tasks

We select the model for the specific domain: for portraits — a pair of Real-ESRGAN + GFPGAN, for architecture — pure Real-ESRGAN, for anime/art — a specialized version with anime weights. We wrap everything in an API service that easily integrates into your pipeline. We use tiled inference to process images of any size without OOM.

How to set up an upscaling pipeline

1. Install dependencies: pip install basicsr realesrgan gfpgan.
2. Download pretrained weights Real-ESRGAN_x4plus.pth and GFPGANv1.4.pth.
3. Run inference on a single image: use the example code below for testing. Then scale to batch with DataLoader.

Real-ESRGAN — practical standard

import torch
import numpy as np
from PIL import Image
from basicsr.archs.rrdbnet_arch import RRDBNet
from realesrgan import RealESRGANer

def upscale_image(
    image_path: str,
    scale: int = 4,
    model_name: str = 'RealESRGAN_x4plus',  # or 'RealESRGAN_x4plus_anime_6B'
    tile_size: int = 512,    # for large images — tile processing
    half_precision: bool = True
) -> np.ndarray:
    """
    tile_size=512 for 6GB VRAM, tile_size=0 (whole image) for 24GB VRAM.
    half=True — FP16, saves ~50% VRAM.
    """
    model = RRDBNet(
        num_in_ch=3, num_out_ch=3,
        num_feat=64, num_block=23, num_grow_ch=32,
        scale=scale
    )
    upsampler = RealESRGANer(
        scale=scale,
        model_path=f'weights/{model_name}.pth',
        model=model,
        tile=tile_size,
        tile_pad=10,      # tile overlap for seamless stitching
        pre_pad=0,
        half=half_precision,
        device='cuda'
    )

    img = np.array(Image.open(image_path).convert('RGB'))
    output, _ = upsampler.enhance(img, outscale=scale)
    return output

GFPGAN for face restoration

Real-ESRGAN sometimes creates facial artifacts on portraits. GFPGAN adds face restoration on top of SR:

from gfpgan import GFPGANer

def restore_face_photo(
    degraded_image: np.ndarray,
    upscale: int = 2,
    arch: str = 'clean',         # 'clean' | 'RestoreFormer'
    channel_multiplier: int = 2,
    weight: float = 0.5          # 0= pure GFPGAN, 1= no face enhancement
) -> np.ndarray:
    """
    weight=0.5 — compromise between restoration and preserving individual features.
    At weight=0 faces become 'glossy'.
    """
    restorer = GFPGANer(
        model_path='weights/GFPGANv1.4.pth',
        upscale=upscale,
        arch=arch,
        channel_multiplier=channel_multiplier,
        bg_upsampler=None   # RealESRGANer can be passed for background
    )

    _, _, restored = restorer.enhance(
        degraded_image,
        has_aligned=False,
        only_center_face=False,
        paste_back=True,
        weight=weight
    )
    return restored

Why Real-ESRGAN is the industry standard

The model is trained on realistic data with synthetic degradations (noise, blur, compression), so it works well with real photos. Combining with GFPGAN for faces produces detailed results without artifacts. Our experience shows that for 90% of commercial tasks, this pair is optimal in terms of quality/speed. Furthermore, Wang et al., "Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data" confirms its effectiveness on benchmarks.

Metrics and model comparison

PSNR is not the only criterion: human perception correlates with LPIPS (perceptual loss). Real-ESRGAN, despite a lower PSNR than SwinIR, often looks better subjectively due to higher frequency details.

Batch processing large volumes

from pathlib import Path
import torch
from torch.utils.data import DataLoader, Dataset
from torchvision import transforms

class ImageDataset(Dataset):
    def __init__(self, image_paths: list[str], size: int = 256):
        self.paths = image_paths
        self.transform = transforms.Compose([
            transforms.Resize((size, size)),
            transforms.ToTensor()
        ])

    def __len__(self): return len(self.paths)

    def __getitem__(self, idx):
        img = Image.open(self.paths[idx]).convert('RGB')
        return self.transform(img), self.paths[idx]

def batch_upscale_pipeline(
    input_dir: str,
    output_dir: str,
    batch_size: int = 4,   # for 12GB VRAM and tile_size=0
    scale: int = 4
):
    paths = list(Path(input_dir).glob('*.{jpg,jpeg,png}'))
    Path(output_dir).mkdir(exist_ok=True)

    # For batch inference we use direct forward
    # (RealESRGANer does not support batches, need direct model call)
    model = RRDBNet(
        num_in_ch=3, num_out_ch=3,
        num_feat=64, num_block=23, num_grow_ch=32, scale=scale
    )
    model.load_state_dict(
        torch.load(f'weights/RealESRGAN_x4plus.pth')['params_ema']
    )
    model.eval().cuda().half()

    for path in paths:
        with torch.no_grad(), torch.cuda.amp.autocast():
            img_t = transforms.ToTensor()(
                Image.open(path).convert('RGB')
            ).unsqueeze(0).half().cuda()
            out = model(img_t).squeeze(0).float().cpu()
            out_img = transforms.ToPILImage()(out.clamp(0, 1))
            out_img.save(
                Path(output_dir) / (Path(path).stem + '_4x.png')
            )

Limitations and typical issues

• Texture hallucinations — Real-ESRGAN may add non-existent text on signs. In forensic applications this is unacceptable
• OOM on large images — a 12-megapixel photo at 4x upscale yields 192MP, doesn't fit in memory entirely. Solution: tile_size=512 with tile_pad=10
• JPEG artifacts — blockiness of JPEG artifacts is amplified by SR. Preprocessing: JPEG-aware denoising (nf_denoise from BasicSR)

What's included in turnkey implementation

We provide: a working API on FastAPI with documentation (Swagger), a Docker image for easy deployment, instructions for setting up GPU inference, benchmarking of your data, and one month of support after delivery. Training of the customer's team is available if needed. We guarantee stable operation and optimization for your hardware. The cost of processing one image in batch mode ranges from $0.002 to $0.02 depending on size and model. Order a pilot project to evaluate the quality improvement on your data. Get a consultation — contact us.

Timelines

Budget savings when choosing a ready-made model over development from scratch can reach 4–6 times. We will evaluate your project for free — contact us. We have 5+ years of experience in computer vision, dozens of successful integrations.

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.