Unlike computers, humans process images easier and faster, than text. As users, we become less patient and expect technology to read our minds, and serve the right content. As developers, we have more and more opportunities to work with Machine Learning and Deep Learning. E-commerce companies recognised these trends quickly, introducing image recognition tools, because it brings higher conversion rate. Let’s take a deeper look at visual search: how and where it works, who is using it and how you can do it yourself.
5. • brain process images faster,
like 60 K faster, than text
• image is international
Humans think
with images
6. • stop words for every language
• translated names and descriptions
• handle bad images
To be
or not
to be:
that is
the
question
:
International search
13. viewers are more likely to
purchase a product after
watching a product video
of information received
by the brain is visual
of millennials want
visual search over any
other new technology
respondents preferred
visual information
(except electronics and spirits)
90% 85% 62% 50%
WHY YOU SHOULD CARE?
16. 1. Prepare tags/concepts
(or use ready ones)
2. Prepare the training data
3. Train the model
4. Upload the image
5. Extract tags from the image
Image 2 text
24. • tag each image
with image recognition
• save tags is search engine (f.e. Elastic)
• tag uploaded image and search for tags
• add context
Use image tags
in text search
25. 1. Use a good model
(or build your own if you have years)
2. Extract features from images
in dataset
3. Upload the image
4. Extract feature from the image
5. Look for the most similar features
Image 2 image
26. Simple Image Search Engine from Yusuke Matsui
https://github.com/matsui528/sis
He uses CNN VGG16
trained on the ImageNet*
27. • dataset of 15 M images
• 1000 categories
• collected by people
• organises challenges
• sets trends
28. VGG16 - the winner of
ImageNet Large Scale Visual Recognition Challenge 2014
Up to 90% of accuracy
42. • diverse pictures from each
category and subcategory
• grouping by department
• response time
Training data size
43. LET’S RECAP
1. Visual search blooms now and
will get more popular
2. You can do it yourself, use ready
made tools or make a combo
3. You need good quality images
and large training set