This project demonstrates the integration of the Fast Super-Resolution Convolutional Neural Network (FSRCNN) model onto the Atlas 200 DK platform for real-time facial image enhancement. Combining real-time face detection with efficient image upscaling, the system showcases the potential of deploying advanced AI models on edge devices for low-latency, high-quality visual enhancement.
The Fast Super-Resolution CNN (FSRCNN) is a real-time super-resolution model that builds upon and optimizes the earlier SRCNN architecture.
| Feature | SRCNN | FSRCNN |
|---|---|---|
| Upsampling Technique | Bicubic interpolation | Learnable deconvolution |
| Inference Speed | ~1.6 FPS | ~24 FPS |
| Application | Offline (slow) | Real-time, edge-ready |
Key Enhancements in FSRCNN:
-
Replaces slow bicubic interpolation with efficient deconvolution layers
-
Reduces model complexity and improves speed significantly
-
Ideal for real-time applications with constrained hardware (e.g., Atlas 200 DK)
To run the FSRCNN model on the Ascend 310 chip, it must be converted to a compatible Da Vinci format.
-
MindStudio GUI: User-friendly interface for model configuration and conversion
-
OMG CLI (Offline Model Generator): Powerful command-line tool for advanced model optimization and deployment
Note: CLI is recommended for greater customization and repeatability in automation pipelines.
The Atlas 200 AI Developer Kit is powered by Huawei’s Ascend 310 processor and supports real-time AI inference. Features include:
-
Direct camera input for live face detection
-
Presenter server for web-based result visualization
-
Support for deploying multiple models simultaneously for pipeline tasks
Example image or frame showing successful face detection on live input.
High-resolution output image showcasing improved facial clarity after FSRCNN processing.
