Search
  • Kanav Khurana

The Daily Dispatch #6, Let's clean this up, Part 2

Einstein Vision includes these APIs:

  • Einstein Image Classification—Train deep learning models to recognize and classify images at scale.

  • Einstein Object Detection—Train models to recognize and count multiple distinct objects within an image, providing granular details like the size and location of each object.

  • Einstein OCR (Optical Character Recognition)—Use OCR models to detect alphanumeric text in an image or PDF.

It seems I will need the first two.

Image Classification: To work off of a sufficiently large store of sample images - of plastic bottles, furniture, cans, cigarette buds, etc.


Einstein Object Detection: From a pile of trash, be able to identify the distinct items


Image Classification

You can use pre-trained classifiers or train your own custom classifiers to solve unique use cases.


Image Classification works off of a combination of machine learning and deep learning.


I worked on re-factoring code for a very interesting application back in my University days. We ended up publishing a paper here.

Anyway, I remember a nice example to explain the different terms to you.


Imagine a conveyer belt which has a camera attached to it. Its function is to automatically characterise the different types of fish that are put inside it.


How can you train a machine learning algorithm for this classification?


Well, simple.

Step 1: Training a model

Say, you would like the machine to identify 3 different types of fish - salmon, mackerel and atlantic bluefin tuna.


You put a salmon in the belt, let the camera analyse its different features and also tell the camera that this is a salmon.


Then, you put a mackerel in the belt, let the camera analyse its features and tell the camera that this is a mackerel.


Then, follow the same thing with the tuna.


You expect the camera to analyse the three different types of fish and create an internal picture of a generic salmon, a generic mackerel and a generic tuna.


This process is called training a machine learning model.


Step 2: Testing a model


Next, you expect to have trained the model enough with enough test data that now, when you put in a tuna, say, you now expect the camera to tell you that it is a tuna.


Hopefully that was a simple explanation.


Let’s come back to Salesforce Einstein. Next, we continue with some setup steps:



Clicking preview on the VF page shows us an image of a frog and the result.

So, the stock code works!




As you can see, the model identifies 5 possibilities as to what this image could be, but the highest probability , with 82.6% - is a tree frog.


3 views0 comments

Recent Posts

See All