Image Recognition For Non-Experts
YOW! Data
19 Sept 2017
Gareth Jones - @nomiddlename
What I'm going to talk about
- Serving Ads
- POC (Cloud Vision API)
- POC (Tensorflow)
- Summary
- Dance Number
Serving Ads
Ad Breaches
- Untargeted ads
- Algorithmic selection
- Third-party Ad Networks
- Human error
Now
- Complaint
- Panic
- Fixed
What we want
- Detection
- Fixed
Why Google Cloud Vision?
- Image Recognition is just not our bag, man.
- Managed service
- OCR, Logo detection
- Objects, faces, landmarks
- Inappropriate content
Example
Face Detection
Labels
Logos
Web
Text
Inappropriate Content
Proof Of Concept
Keywords
Solution
Make our own image classifier!
How hard can it be?
Too Hard
It'll take weeks and millions of images
Steal* Someone Else's Work!
*Make use of work freely given for the advancement of humanity
(transfer learning)
Tensorflow
- Machine learning toolkit
- Open-source
- Has pre-trained models
Inception V3
How to do it
- Download TF source
- Compile
- Put images into folders named by category
- Run the tutorial code
- That's it.
Simpler Diagram
Images
- Sports - 170
- Not Sports - 230
Sports
Do the thing
- Took about 10 minutes
- Accuracy = 76.1%
Results
sports = 0.96
not-sports = 0.03
sports = 0.93
not-sports = 0.07
sports = 0.45
not-sports = 0.55
sports = 0.03
not-sports = 0.97
Is it over yet?
- Cloud Vision API
- Easy
- Managed
- Always improving
- Rolling your own image classifier
