1 Machine Learning for Images
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
1
Machine Learning for Images
2
Huge amountof user-generated
content
NOT searchable
NOT monetizable
It’s a big, big image world
3
Image is a matrix of pixels (raster data)
What is specific: image recognition
4
Recognition is sensitive to: lighting
contrast
saturation
blur
noise
What is specific: image recognition
5
sensitive to
geometric transformations (scaling, translation, rotation)
occlusion
hard tofind optimal set of filters
What is specific: image recognition
6
pixels lots of data themselvesin spatial relationships
multiple levels and scales of interestfrom low-level features such as texture to high-level features such as composition
needs data-augmentationcompensate for sensitivity - training with blurred, cropped, scaled, noised, etc
What is specific: ML for images
7
huge architecture (both deep and wide) - requires massive amount of memory and processing power
inter- and intra-class varietyneed to describe the universe (huge and diverse datasets required to feed the data greedy CNNs)
takes time10+ days for large architectures, even after 10x reduction thanks to using GPUs
What is specific: ML for images
8
cuda-convnet python interface, fermi-generation nVidia GPU, no multi-GPU support
cuda-convnet2an upgrade to cuda-convnet, optimized for new kepler-generation nVidia GPUs, multi-GPU support
caffedeep learning framework, developed by Berkeley Vision and Learning Center, big community of contributors
Convnet implementations for images
9
torch7 ML algorithms, CNN extensions: fbcnn by Facebook, used by Google DeepMind
theanopython library, open-ended in terms of network architecture & transfer functions
…many others
Convnet implementations for images
10
Imagga auto tagging
11
CNN classifier with ~3000 everyday-life objects
Imagga Auto Tagging
12
Expansion with semantically related concepts
Imagga Auto Tagging
business meeting presentation
laptop
desk
people
13
semantic expansion ‘car’ -> ‘vehicle’ -> ‘mean of transportation’
feedback-loopinstant learning from user feedback, to be released in May
custom trainingwith specific set of tags
Imagga Auto Tagging
14
What can be builtwith Image Recognition
15
Personal Photo Applications
•Apps for mobile photos organization• Integration in telecom solutions•Cloud services for consumers•Device manufacturers
http://getsliki.com
16
Integration with analytics platforms
17
Image Processing Services
19
Big image data management and organization Image driven platforms (DAMs)
Contextual advertisingInteractive/behavioural campaigns (adSense like)
User profiling insight market, profiling based on image content
Interactive campaigns for brands new ways to interact with customers
Use Cases
20
Imagga APIs
21
Imagga API: How it works
22
Sign Upwww.imagga.com
Free Hacker Planto try out our image tagging API
Developer Planany of the APIs & within larger limits
23
ML Internship learn the AI magic
work with machine learningdata collectioninternal tool-chains
email us: [email protected]
Related Documents
