IBM Watson Visual Recognition
976 words · 5 minutes
What is IBM Watson?
If you've never heard of Watson, this service is a suite of enterprise-ready AI services, applications, and tooling provided by IBM. Watson contains quite a few useful tools for data scientists and students, including the subject of this post today: visual recognition.
If you'd like to view the official documentation for the Visual Recognition API, visit the API Docs.
To be able to use Watson Visual Recognition, you'll need the following:
- Create a free account on IBM Watson Studio.
- Add the Watson Visual Recognition service to your IBM Watson account.
- Get your API key and URL. To do this, first go to the profile dashboard for your IBM account and click on the Watson Visual Recognition service you created. This will be listed in the section titled Your services. Then click the Credentials tab and open the Auto-generated credentials dropdown. Copy your API key and URL so that you can use them in the Python script later.
- [Optional] While not required, you can also create the Jupyter Notebook for this project right inside Watson Studio. Watson Studio will save your notebooks inside an organized project and allow you to use their other integrated products, such as storage containers, AI models, documentation, external sharing, etc.
Calling the IBM Watson Visual Recognition API
Okay, now let's get started.
To begin, we need to install the proper Python package for IBM Watson.
Next, we need to specify the API key, version, and URL given to us when we created the Watson Visual Recognition service.
= = =
Now, let's import the necessary libraries and authenticate our service.
[Optional] If you'd like to tell the API not to use any data to improve their products, set the following header.
Now we have our API all set and ready to go. For this example, I'm going to include a
dict of photos to load as we test out the API.
Now that we've set up our libraries and have the photos ready, let's create a loop to call the API for each image. The code below shows a loop that calls the URL of each image and sends it to the API, requesting results with at least 60% confidence. The results are output to the console with dotted lines separating each section.
In the case of an API error, the codes and explanations are output to the console.
= = = =
Here we can see the full result set of our function above. If you view each of the URLs that we sent to the API, you'll be able to see that it was remarkably accurate. To be fair, these are clear high-resolution, clear photos shot with a professional camera. In reality, you will most likely be processing images that are lower quality and may have a lot of noise in the photo.
However, we can clearly see the benefit of being able to call this API instead of attempting to write our own image recognition function. Each of the classifications returned was a fair description of the image.
If you wanted to restrict the results to those that are at least 90% confident or greater, you would simply adjust the
threshold in the
When your program runs, it should show the output below for each photo you provide.
---------------------------------------------------------------- Image Title: Grizzly Bear Image URL: https://example.com/photos/image1.jpg brown bear ( 0.944 ) bear ( 1 ) carnivore ( 1 ) mammal ( 1 ) animal ( 1 ) Alaskan brown bear ( 0.759 ) greenishness color ( 0.975 ) ----------------------------------------------------------------
Now, this was a very minimal implementation of the API. We simply supplied some images and looked to see how accurate the results were. However, you could implement this type of API into many machine learning (ML) models.
For example, you could be working for a company that scans their warehouses or inventory using drones. Would you want to pay employees to sit there and watch drone footage all day in order to identify or count things in the video? Probably not. Instead, you could use a classification system similar to this one in order to train your machine learning model to correctly identify items that the drones show through video. More specifically, you could have your machine learning model watch a drone fly over a field of sheep in order to count how many sheep are living in that field.
There are many ways to implement machine learning functionality, but hopefully this post helped inspire some deeper thought about the tools that can help propel us further into the future of machine learning and AI.