Image classification is one of
the most widely used areas of machine learning.
IBM Watson provides industry leading services,
so that you can create models that can
identify objects in an image.
But first, how can we use
computers to identify what is in an image.
Let's start with a simple example of a picture of a dog.
If you showed this picture to someone and then
asked them to tell you what was in this picture,
what would you expect them to say?
Well, you would expect a dog or
perhaps a Golden Retriever which is a breed of dog.
They might also say animal, four-legged animal,
or something about the color of the dog's fur.
We could call these descriptors image labels
or key words that describe the object in the photo.
Image classification does the same thing.
We provide a photo to a machine,
and the machine will provide us with labels
and key words describing the photo.
Watson Visual recognition is an industry leading service
within IBM Watson that performs image classification.
When you provide IBM Watson with an image,
IBM Watson provides you with a list of
labels describing the image
along with its confidence scores.
These confidence scores are between zero to one and
describes how confident that
the label describes the image,
with zero being extremely unconfident,
and one being extremely confident.
The label Retriever dog has a confidence score of 0.99.
Which means that IBM Watson is
99 percent confidence that
there is a Retriever dog in the picture.
While the labeled beige color
has a confidence score of 0.75,
which means that IBM Watson is 75 percent
confident that the picture contains the color beige.
One of the biggest advantage of
IBM Watson visual recognition
is that it is extremely user-friendly,
because the models used for
image classification was already
pretrained and prepackaged for you to use out of the box.
All you have to do is provide the images you want to
perform image classification on to IBM Watson.
IBM Watson will return a list of
labels along with their competence scores.
This makes computer vision extremely easy
to use for someone that is new to machine learning.
Here are some more examples; a bike helmet,
an emergency ambulance, and a music concert.
You can see it does a pretty good job
identifying the actual labels if not relevant keywords.
In the lab exercise,
you'll get a chance to try out
this visual recognition service.
You could even start thinking of what kind of
images you might want to upload to try to classify.
Just an important note about confidence scores.
They range between zero and one for each of the labels,
and as such, the confidence scores
across all labels don't need to add up to one.
Based on data used to train IBM Watson,
the confidence scores represents how confident the model
believes it has identified
a particular label in the image.
But note that the scores are not numbers you
can say are some percentage higher than others.
You cannot say that a confidence score of one
is two times the confidence score of 0.5.
For example, if in one image, image A,
the label cat has a confidence score of one.
In another image, image B,
the label cat has a confidence score of only 0.5.
Then you cannot say that a cat is two times likely
to appear in image A than in image B.
But you can say, that it's much more likely that
there's a cat in image A than in image B,
according to the classifier.
For those who are more mathematically inclined,
confidence scores are created using
a sigmoid normalization function.
So a score of one actually
represents more than two times the confidence than 0.5.
Finally, do note that labels with
a confidence score lower than 0.5 by default,
will not show up in the results.
Okay. What about this image?
You can see it's a plate of Sushi or
Sushi rolls which is a popular Japanese dish.
But when you send it to IBM Watson,
it doesn't seem to do a very good job
at recognizing it as Sushi.
You see what we've been doing so far is
IBM Watson's general classifier for visual recognition.
Which as you can imagine is good
across the board at identifying most objects.
This is what we can call more breadth and less depth,
because it's pretty good at classifying across
a variety of objects or scenes but
doesn't do a very good job at
very specific objects or scenes like a plate of Sushi.
But in addition to the general classifier,
IBM Watson also comes with a number of
other classifiers that you can use and try.
In fact, there's a classifier just for food items,
which is very good at classifying images of
food and actually only food.
So we can say that this classifier
has a lot of depth for food,
but not a lot of breadth across other categories.
That's because of the images and labels that were
used to train these classifiers.
Food images and food labels for the food classifier,
and more generic images and
generic labels for the general classifier.
So using this food classifier,
we get a highly confident label of Sushi
and a somewhat confident label of California roll.
So if you're interested in
classifying between different food dishes,
I recommend trying out this food classifier.
There's also the explicit classifier,
which evaluates whether an image might be pornographic.
It returns confidence scores for
explicit and not explicit.
While you may or may not see use cases
for the explicit classifier in your particular industry,
some companies that have users uploading images
may sometimes need a way to filter out or
sensor potentially pornographic images
without humans needing to look
through all the images that were uploaded.
In this video, we saw how image classification works with
Watson Visual Recognition Service.
Thank you for watching.
