Hack4 fi tds presentation 07022016 lowresolution

Hack4FI Similar Images Search Engine

Team: Timo and Ouz from Top Data Science & Pia from YLE

7.2.2016

Automatically recognize similar images using machine learning

• State-of-the-art AI algorithms (deep learning)• Training on ImageNet (~1 Million images)– 2 weeks on GPU

• Hack4FI image sets used to fine-tune the system (~ 20 hours on GPU)

• Purely based on pixel values, no meta data

Data sources:

• 1500 images• Sources:– YLE archives– Satakunnan museo– Tallinn City Museum– Suomi syö ja juo– Gallen Kallelan museo– Svenska LIteratursällskapet– Valokuvataiteen museo– Aalto University commons

Input image (Google images)

Closest match (YLE archives)

Input image (Google Images)

Closest match (Suomi syö ja juo)

Input image (Google images)

Closest match (Source: Svenska litteratursällskapet)

Input image (Google Images) Closest match (Source: Valokuvataiteen

museo I.K Inha )

Input image (Google Images)

Closest match (Source: Svenska litteratursällskapet)

Use cases & scenarios…Object recognition and automatic annotation/tagging

Use cases & scenarios…Count the number of persons in images

Use cases & scenarios…Estimate the age of persons in images

Use cases & scenarios…Map/Group similar images with each other (clustering)

Top Data Scientist at your service…

• Image Analysis & Understanding• Data Analytics• Software Engineering

[email protected]@topdatascience.com

http://topdatascience.com/