Super Resolution with CoreML

on YouTube ↗24m

" summary

Implementation shows on‑device super resolution with CoreML and Swift that keeps image clarity while shrinking transfer size. Training uses Python + Keras, converted with coremltools; a produced model is small at 4,540 KB model. Baseline SRCNN has ~80–85k parameters and runs ~1.5s on iPhone X; FSRCNN reduces to ~12,000 parameters and runs ~1.1s (~40–45% faster) with comparable PSNR. A CocoaPods SuperResolutionKit demo works but needs task‑specific retraining and is not production‑ready for non‑manga images.

AI-written summary · about

" chapters (19)

1. 0:00Start
2. 1:55Resolution vs. File Size
3. 3:57Super Resolution with CoreML
4. 6:46Kernel and Convolution
5. 8:10Convolutional Neural Network (CNN)
6. 8:30Super Resolution with CNN (SRCNN)
7. 9:54Hard to collect training data?
8. 10:20Easy to collect training data for SRCNN
9. 11:05Training SRCNN
10. 13:16SRCNN Model Structure 85,000 parameters
11. 13:30Training code
12. 14:15CoreML Model file size
13. 17:30Performance
14. 19:26Limitation
15. 20:18Motivation for ML in Mobile Apps
16. 21:24Choose the best way
17. 21:56Should I learn Python?
18. 22:44Dynamic Member Lookup
19. 23:18Recap

" related talks

• 07mSuper Resolution with CoreMLKentaro Matsumae · try! Swift Tokyo 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
• 31mPower-Up Your App With Apple’s Machine Learning FrameworksKim Foong Chow · iOS Conf SG 2017
• 34mThe Potential of IMAGE Processing in Swift for DevelopersElina Semenko · Swift Heroes 2023
• 01h 03mTraining dynamic ML models on iOS 15Martin Mitrevski · Mobius 2021 Moscow
• 27mMachine Learning and MobileBrett Koonce · Scale By The Bay 2019