Low to High Resolution
summary
Keras implementation trains a roughly 120‑layer dense residual network on DIV2K (~800 images, 16‑patch batches, ~86 epochs on a p2.xlarge) and reports baseline PSNR ≈30 for bicubic, modest PSNR gains with MSE but perceptual drops (down to ~26) when using VGG19 perceptual loss; adding a GAN discriminator increases realism but is unstable; training on JPEG‑compressed inputs reduces artifacts at the cost of fine detail; model‑weight averaging across two models yields a controllable denoising tradeoff; code is published and pip‑installable as a Keras implementation.
AI-written summary · about
chapters (11)
- 0:00Super resolution project intro
- 2:07Neural network super resolution basics
- 8:30Evaluation metrics and training
- 12:06Model limitations and failures
- 14:37JPEG artifacts and noise
- 17:51Perceptual loss for quality
- 22:03Adversarial networks for realism
- 26:19Denoising vs adversarial results
- 28:50VGG feature layers for training
- 30:16Model averaging for denoising
- 31:21Optimizing prediction time
related talks
12mENHANCE! Upscaling Images with Neural Networks
26mHit-reco: ProtoDUNE Raw Data Denoising with DL models
56mTutorial on Deep Learning II
01h 30mDeep Learning Techniques for Monocular Depth Estimation
40mVisual concept learning from few images
01h 17mDeep generative models for image...
44mGenerative Adversarial Networks (GANs) and Stable Diffusion
52mDeep Generative Models And Unsupervised Methods For Inverse Problems