Getting Started with Orange 14 Image Analytics - Clustering Video Lecture

Data can come in all shapes and sizes.
It can be a simple spreadsheet or a complex text,
audio, video or even an image.
Today, I will show you how Orange transforms images into numbers
and in this way enables machine learning.
First, we will have to install Image Analytics add-on.
Go to Options,
Add-ons,
and install Image Analytics.
Restart Orange for the add-on to appear.
Now, let's start with an example.
I have collected 19 images of domestic animals
and placed them all in the same folder.
I can load them in Orange with Import Images.
Open the widget and select the folder.
Now, connect Image Viewer to Import Images.
Looks like we have all the animals here.
Import Images actually constructs and outputs a data table.
But how does this data table look like?
Let's check this in a Data Table widget.
All I have is a bunch of meta information.
Image name, image file path,
size of the file, image width and height in pixels...
Nothing that will help me with machine learning.
We need image descriptors
that is numbers which describe the content of these images.
We will transform raw images into their vector representation
using a deep neural network
which was trained on millions of real-life images.
Retrieved vectors are also called image embeddings.
We embed our images into a multi-dimensional feature space.
Connect Import Images to Image Embeddings.
The widget sends images to the server
and computes embeddings remotely.
Let's see what we got back.
Now, we don't only have meta information,
but also 2048 additional features to profile the images.
Great! With embeddings we can compare the images
and compute their similarity.
Let us pass the data to Distances
and select Cosine distance
as it usually works best for images.
Then, we can pass the distance matrix to Hierarchical Clustering
and observe the dendrogram.
Seems like embeddings indeed makes sense.
All the chickens are grouped together
and all the cows, too.
I can even check each cluster in an Image Viewer.
Simply wonderful.
But I know what you're going to say.
You think I'm trying to trick you,
that I intentionally selected similar photos,
so they would get clustered together.
Let's double-check our procedure
with the most famous cow in Europe - the Milka cow.
This image is quite different from the others.
It is a frontal photograph with digitally enhanced colors.
I have retrieved the image from the internet
and added it to my domestic animals folder.
If I press reload in Import Images
the new image gets loaded
and the whole workflow is instantly updated.
Let us again open Hierarchical Clustering
and Image Viewer.
Seems like Milka is correctly clustered with its kind.
Today we've learned how to work with images in Orange,
profile them using deep model embeddings,
and perform clustering.
In the next video, we will go a step further.
We will be building predictive models for image classification.