Super Resolution with CoreML

on YouTube ↗07m

" summary

Implementation of super resolution with CoreML converting low-res manga to high-res runs fully on-device, exporting a custom-trained SRCNN model that restores line clarity and screen-tone texture lost by generic anime models. Training uses Python and chainer on a manga dataset, producing a ~400 KB .mlmodel that generates a 2× upscaled image in about one second on an iPhone 7. Code and training scripts are open-sourced as the SR CNN kid pod for Swift apps.

AI-written summary · about

" chapters (9)

1. 0:00Start
2. 0:37Super Resolution (SR)
3. 0:48SR Method
4. 2:11Overview
5. 2:38How to get MLModel
6. 3:12#231. Illegible Serif characters
7. 3:56Training Environment
8. 4:53Run Super Resolution process
9. 6:20Recap

" related talks

• 24mSuper Resolution with CoreMLKentaro Matsumae · try! Swift NYC 2018
• 24mSuper Resolution: The scaler of tomorrow, here today!Nick Chadwick · Demuxed 2019
• 40mDeep Learning on iOSShuichi Tsutsumi · 360iDev 2017
• 30mLearning CoreML with funAndrea Baldon · #Pragma Conference 2019
• 40mNavigating the Machine Learning Landscape for Apple DevicesMeghan Kane · Swift & Fika 2018
• 10mMachine Learning And React NativeRay Deck · Chain React 2018
• 45mMobiconf 2018: Meghan Kane, NovodaMeghan Kane · Mobiconf 2018
• 34mDemystifying ML in iOSSanfilippo, Malesani · Swift Heroes 2018