Image Enhancement Web Service

Overview

A professional-grade AI image enhancement service that transforms mobile phone photos into high-quality, professional images. Built with state-of-the-art deep learning models, the service offers a modular pipeline combining super-resolution, face restoration, low-light enhancement, and color correction.

Optimized for deployment on HuggingFace Spaces with free GPU access, making high-quality image enhancement accessible to everyone.

Key Features

Super-Resolution (Upscaling)

• Real-ESRGAN: 4x upscaling with preserved details
• super-image: Alternative upscaling with tile processing for large images
• Supports both 16x (for tiny images) and 4x (standard) upscaling factors

Low-Light Enhancement

• CLAHE: Contrast Limited Adaptive Histogram Equalization
• CSEC: Color and Spatial Enhancement for low-light conditions
• Brightens dark images while reducing noise

Color Correction

• Deep White Balance: Learning-based white balancing
• Exposure Correction: Automatic brightness and contrast adjustment
• Fixes color casts and underexposed photos

Face Enhancement

• GFPGAN: Generative Facial Priors GAN for face restoration
• Reconstructs facial details in low-quality or blurred faces
• Maintains natural appearance without over-processing

Smart Analysis

Automatic detection of required enhancements:

• Brightness check → applies low-light enhancement if too dark
• Face detection → applies GFPGAN if faces found
• Resolution check → chooses appropriate upscaling factor

Algorithm Performance (GTX 1650 Ti)

Tested on 1728×2304 test image (≈4MP):

| Algorithm | Time | GPU | Notes | |-----------|------|-----|--------| | Real-ESRGAN (4x) | 99.8s | ✅ | High quality, 4× upscaling | | GFPGAN (face) | 3.0s | ✅ | Face restoration | | CLAHE | 162ms | ❌ | CPU-based | | White Balance | 64ms | ❌ | CPU-based | | Exposure Correction | 58ms | ❌ | CPU-based |

Total pipeline time: ~105 seconds for full enhancement (upscale + face + color + exposure)

API Endpoints

POST /api/v1/enhance
  Request: { image: base64, algorithms: ['auto'], parameters: { upscale_factor: 'auto', quality: 'high' } }
  Response: { enhanced_image: base64, metadata: {...} }

POST /api/v1/enhance/batch
  Request: Multiple images
  Response: Array of enhanced results

GET /api/v1/algorithms
  Response: List of available algorithms with descriptions

GET /api/v1/status
  Response: System health, GPU status, model load times

Tech Stack

• Backend: Python, FastAPI, Uvicorn
• AI Models: PyTorch, Real-ESRGAN, GFPGAN, OpenCV
• Hardware: NVIDIA GPU (CUDA 11.8+)
• Deployment: Docker, HuggingFace Spaces, Railway, Render
• Storage: Google Drive API (batch processing), Local FS

Validation Results

✅ 15/15 images successfully enhanced (100% success rate)

• Proper 4x upscaling verified: 1728×2304 → 6912×9216
• Face enhancement maintained natural appearance
• Low-light images brightened without noise amplification
• Color corrections eliminated color casts

User feedback: "I find the result good this time" — after iterative improvements.

Architecture

User Upload → Image Analyzer → Enhancement Pipeline → Output
     ↓              ↓                  ↓             ↓
  Base64/File  Brightness/Face/  Real-ESRGAN →  Enhanced
                 Resolution        GFPGAN       Image
                               CLAHE/WB/Exposure

Pipeline Flow

1. Analyze image properties (brightness, faces, resolution)
2. Decide which algorithms to apply (auto mode)
3. Execute enhancement pipeline sequentially
4. Return enhanced image + metadata

Challenges & Solutions

Challenge 1: GPU Memory (4GB VRAM limit)

Large images caused CUDA out-of-memory errors. Solution: Implemented tiling for large images, reduced batch size, processed sequentially instead of parallel.

Challenge 2: Model Compatibility

Different models expected different input formats and preprocessing. Solution: Created unified BaseEnhancer interface; each algorithm handles its own preprocessing, returns standard format.

Challenge 3: JSON Serialization

NumPy arrays cannot be directly JSON-serialized. Solution: Convert to PIL Image → base64 string for API responses; also optionally save to disk.

Challenge 4: Performance

Combined pipeline could take several minutes. Solution: Made pipeline asynchronous with progress tracking; user gets task ID and polls for completion.

Deployment Options

HuggingFace Spaces (RECOMMENDED)

• Free GPU (A10G/T4), 16GB RAM, 50GB storage
• Automatic builds from GitHub
• 1-click deployment

Railway

• GPU-enabled instances
• Easy config via environment variables

Render

• Web Service with GPU
• Persistent storage for models

Local Docker

docker build -t img-enhance .
docker run -p 8000:8000 img-enhance

Future Enhancements

• Video enhancement (frame-by-frame)
• Batch processing from cloud storage (S3, GCS)
• User accounts and enhancement history
• Subscription-based credit system
• REST + WebSocket for real-time progress
• Mobile SDK (iOS/Android)

Try It Out

Since deployment requires GPU, the easiest way is to deploy to HuggingFace Spaces — follow the setup guide in the repo's README.

Once deployed, send a POST request with a base64-encoded image and receive the enhanced version in seconds.