Image similarity / cheat sheet
Use this cheat sheet
If you build projects where you want to find similar images.
What is image similarity?
Image similarity is a way to quantify how similar 2 images are.
You do this by converting all images in your dataset to compressed vectors with a deep learning model.
These vectors are added to an index.
When you want to find a similar image in your dataset to a new image, you run the new image through the model as well. Then you compare the new image’s vector to all dataset images in the index to find the most similar ones.
Image similarity search is fast since the data becomes so compressed.
Example use cases
Search and find similar looking items.
Search and find items that are difficult to filter using keywords.
Nothing special needs to be done for this problem type when you add your image dataset to the platform.
Select one of the different ways in the Datasets view; Import from your data warehouse, directly from your local computer, via data API, or from a URL.
Use the Experiment wizard and in the Problem type tab select choose Image similarity.
This will build a model that will show up on the Modeling canvas.
Navigate to the Settings tab and confirm that it’s only 1 epoch. That’s all you need.
Click Run to start the training.
Skip the Evaluation view. Since you only train for 1 epoch.
Create new deployment
In the Deployment view, click New deployment and select Similarity search.
Select your experiment, epoch 1 for deployment, and Image embedding as Output feature.
Finally, click Create.
Now all the images in the dataset pass through the model once, and the platform builds the index.
Click Enable to make the deployment available for REST API calls.
Test the deployment
Click Open web app to test the deployment in out web app.
API call workflow
This is what will happen when you send an image to your deployed model.
The new image will pass through the deployed model and be converted into a vector.
Compare the new image’s vector with all vectors in the index.
The model returns the most similar images.