Scaling up Deep Learning Based Super Resolution Algorithms Xiaoyong - PowerPoint PPT Presentation

Scaling up Deep Learning Based Super Resolution Algorithms Xiaoyong Zhu Microsoft Cloud AI Group

CNTK implementation

Let’s Enhance

Image Source

Image Source

because human vision is more sensitive to luminance (“black and white”) differences than chromatic differences

A few milestones including SRCNN, VDSR, DRRN, SRGAN

• SRCNN (First to apply deep learning to Super Resolution, 2014) • VDSR (Very Deep Convolutional Networks, 2015) • DRRN (Deep Recursive Residual Network, CVPR 2017) • SRGAN (Photo-Realistic using GANs, CVPR 2017) • EDSR (Enhanced version using part of SRGAN’s work. Winner of NTIRE2017 Super resolution challenge) • NTIRE Challenge (New Trends in Image Restoration and Enhancement) is a challenge in this area (http://www.vision.ee.ethz.ch/ntire17/)

linear bicubic bilinear Image Source

• Bicubic interpolation • VDSR (Very Deep Convolutional Networks, 2015) • DRRN (Deep Recursive Residual Network, CVPR 2017) • SRGAN (Photo-Realistic using GANs, CVPR 2017) • EDSR (Enhanced version using part of SRGAN’s work. Winner of NTIRE2017 Super resolution challenge) • NTIRE Challenge (New Trends in Image Restoration and Enhancement) is a challenge in this area (http://www.vision.ee.ethz.ch/ntire17/)

http://cs231n.github.io/understanding-cnn/

Bicubic Link to paper SRCNN

• Bicubic interpolation • SRCNN (First to apply deep learning to Super Resolution, 2014) • DRRN (Deep Recursive Residual Network, CVPR 2017) • SRGAN (Photo-Realistic using GANs, CVPR 2017) • EDSR (Enhanced version using part of SRGAN’s work. Winner of NTIRE2017 Super resolution challenge) • NTIRE Challenge (New Trends in Image Restoration and Enhancement) is a challenge in this area (http://www.vision.ee.ethz.ch/ntire17/)

ResNet architecture Image frequency CNTK Code

Code available in CNTK

• Bicubic interpolation • SRCNN (First to apply deep learning to Super Resolution, 2014) • VDSR (Very Deep Convolutional Networks, 2015) • SRGAN (Photo-Realistic using GANs, CVPR 2017) • EDSR (Enhanced version using part of SRGAN’s work. Winner of NTIRE2017 Super resolution challenge) • NTIRE Challenge (New Trends in Image Restoration and Enhancement) is a challenge in this area (http://www.vision.ee.ethz.ch/ntire17/)

The coolest idea in ML in the last twenty years - Yann Lecun

x D D(x) G z G(z) D(G(z)) https://www.slideshare.net/xavigiro/deep-learning-for-computer-vision-generative-models-and-adversarial-training-upc-2016

http://people.eecs.berkeley .edu/~junyanz/projects/gvm/

Image source: http://kvfrans.com/visualizing-features-from-a- convolutional-neural-network/

https://www.cntk.ai/pythondocs/CNTK_302A_Evaluation_of_Pretrained_Su per-resolution_Models.html Bicubic SRGAN DRRN

https://www.cntk.ai/pythondocs/CNTK_302A_Evaluation_of_Pretrained_Su per-resolution_Models.html Bicubic DRRN SRGAN

SRCNN VDSR DRRN SRGAN EDSR

here here

Scalable Machine Learning using Kubernetes • Slides: bit.ly/DLwithK8S • Tutorial for deploying DL with K8S using acs_engine: bit.ly/K8SwithACSEngine • Tutorial for deploying DL with managed K8S: aka.ms/AKS_GPU • Azure Machine Learning simplification to K8S: aka.ms/AMLtoACS • Batch AI for training DL at scale: bit.ly/deepbait