Welcome to our lightweight student model for video restoration!
Built in just one month by a team of 3 students, this project explores Knowledge Distillation using the powerful Restormer architecture as a reference.
While the results aren't perfect yet, this was an incredible learning journey into transformer-based video enhancement. 🚀
🔗 📽️ Final Output Video:
Watch Output on Google Drive
📄 📚 Report Document:
Read Report PDF
🎯 Goal:
To design a compact student model for video sharpening and restoration, learning from a teacher model based on Restormer.
🧠 Technique Used:
- Knowledge Distillation (KD)
- Inspired by Restormer
- Custom student architectures for:
- Defocus deblurring: | SSIM: 0.9216 | PSNR: 28.33 dB
- Motion deblurring: | SSIM: 0.9595 | PSNR: 33.16 dB
- Deraining: | SSIM: 0.9078 | PSNR: 30.52 dB
📆 Duration:
1 Month (Intense)
📂 💻 Code Files:
Check the code above in this repository.
- Tiny student model trained via knowledge distillation
- Inference pipeline for full video processing
- Batch-wise processing for faster testing
- Multiple sub-models stacked: Defocus → Derain → Deblur