Inspection AX›Image Restoration
Image Restoration
SIRNet · SNUAILAB Image Restoration Network
A lightweight deep learning image restoration engine that fixes noise, blur and low resolution in low-dose X-ray and CT images in real time, sharpening the fine structures that battery inspection and medical imaging depend on.
Image Restoration
In precision manufacturing and medical diagnosis,
low-dose imaging matters more than ever
DR (Digital Radiology) is an imaging technique that converts transmitted radiation signals into digital form.
Medical settings and battery inspection processes are divided into Static and Dynamic by imaging method, and each faces its own low-dose noise dilemma.
-
Static DR(one precise shot taken in a fixed position)
When the dose is insufficient or reduced, noise builds up until fine structures can no longer be distinguished, risking diagnostic and judgment errors.
Single exposure→Single image acquisition→Single-view capture→Still image
-
Dynamic DR(video technique capturing many frames continuously in real time during procedures or line runs)
Because exposure is prolonged, imaging must use an ultra-low dose of around 1–1.5K, and the severe quantum mottle noise makes real-time restoration essential.
Continuous exposure→Continuous image acquisition→Real-time frames→High-quality video from continuous frames
Image Restoration
IR Approaches Compared
Our Own Backbone Architecture
| Item | CNN-based model (conventional) | SIRNet (SNUAILAB) |
|---|---|---|
| Inference speed | Slow inference due to a deep network structure | Maximized parallel efficiency enables real-time deployment on site |
| Compute Efficiency | Heavy convolution load drives excessive GPU memory use | Self-Attention-based design greatly reduces GPU memory use |
| Restoration quality | Hard to model long-range dependencies, limiting fine-detail restoration | Global context understanding restores fine details accurately |
- 33.7 dBPSNRCNN 28.1 dB
- 0.927SSIMCNN 0.842
- 12.8 fpsInference speedCNN 3.2 fps
- 4.1 GBGPU memory useCNN 11.2 GB
Quantitative Comparison with Competitors
Top-level quality, with
178× faster inference · 90%+ less GPU memory
An engine that denoises 1.5K × 1.5K high-definition video seamlessly in real time at 30 fps.
- 178×Faster inferenceAverage inference time 0.0845 s per image
- 90%↓GPU memory savedGPU memory use 1,832 – 3,232 MiB
- 30 fpsReal-time denoising1.5K × 1.5K high-definition video
Core Technology
Noise, blur and resolution
restored in a single lightweight model
- Denoising
Denoising
Removes the artificial noise of low-dose imaging without a trace
- Deblurring
Deblurring
Quickly removes afterimages and blur caused by high-speed motion or shaking
- Super Resolution
Super Resolution
Greatly increases resolution with AI and no sensor replacement, switching with denoising in a single model
- AAAI 2026
GuidNoise
Learns real noise patterns from a single reference image and generates large volumes of training data without additional X-ray exposure
Even when equipment or dose conditions change, retaking a single reference image is enough to update right away.
01Feature
Low-dose image
denoising
Removes the artificial noise of low-dose imaging without a trace, achieving state-of-the-art (SOTA) performance of PSNR 33.7 dB.
02Feature
Static · Real-Time
Image Restoration
Dynamic IR restores 1.5K × 1.5K video seamlessly in real time at 30–40 frames per second, enabling high-speed inline 100% inspection.
03Feature
Industrial battery
X-ray restoration
Maximizes visibility of the battery's internal anode structure without cathode distortion or blur, and sharpens the identification of micro-cracks and internal electrode alignment.
04Feature
Medical X-Ray · CT
Image Restoration
Restores fine tissue boundaries even in ultra-low-dose images without creating hallucinations or distortion, securing precise diagnostic quality with no additional exposure.
Why SNUAILAB?

World-class core technology
Our proprietary SIRNet backbone achieves SOTA-level performance in denoising.

Trainable without ground truth
Removes real noise with self-supervised learning, even without paired high-quality originals.

Decisive quantitative advantage
With 178× faster inference and 90% less GPU memory, real-time use in the field is possible.

Proven industry references
Applied in battery inspection (Companies L and S) and medical imaging (Company V).
With high-speed, low-dose Image Restoration AI,
SNUAILAB is transforming digital radiography (DR) inspection.
For sharper image restoration,
start with SNUAILAB SIRNet
Maximize your business values
by adopting industry-specific AI solutions.