Intro: Hands on Azure Custom Vision

Bogusław Błoński
Nov 5 · 3 min read

Context

On daily basics I’m a #Xamarin developer or Full Stack and have nothing common with Machine Learning, nevertheless AI is something I would like to investigate.

During some cafe break I chatted about: https://www.pgs-soft.com/case-studies/creating-image-recognition-software-to-find-cracks-on-smartphones/

Conclusion from that was Machine Learning is hard, requires lot of skills and time.

Ok I’m not an expert on that, but problem can’t be left, lets check other ways.

Other ways = Cloud

You can’t do it by myself, so lets pass it others.

As I’m .NET developer Azure looks as natural choice.

Lucky for me it turned out that my favorite Jim Bennett did video

as well as blogged about and I have proper Visual Studio licencing

( https://azure.microsoft.com/en-us/pricing/details/cognitive-services/custom-vision-service/)

I was very optimistic about it, all looks Super Easy:

  1. Upload image
  2. Tag the images
  3. Train the model
  4. Evaluate the result

Experiment: I’m a broccoli 0,2%

Goal of experiment: teach model to detect broccoli.

My first thought was “how to teach a baby talk”.

Lets show him what is “broccoli ” what is for sure “not broccoli” and what is something similar and than go more and more into details.

Step 1: Uploaded/Tagged broccoli images without background

Step 2: Uploaded/Tagged broccoli images with background

Step 3: Quick Tests

  • Downloaded some random broccoli images from internet, test results ware that they are broccoli, good
  • Model knows what is broccoli , but will it behave as little goose that all living pets are its mother? upload my face image, great and yes I’m broccoli.

Step 4: Tag faces

Step 5: Quick Test: downloaded random faces, faces are not broccoli.

Joker green hair is NOT broccoli.

Step 6: Quick Test: Is something similar to broccoli is broccoli?

As expected The Cauliflower is broccoli.

NEXT Steps: Upload more images and run more tests

Cool option is that you can tag images on tests predicted result.

added Cauliflower so lets see how it will behave on “pink” Cauliflower

Summary

Azure portal is very intuitive. I used only quick training. I didn’t expect that uploading 9 images will deliver any good image classification.

Biggest problem is to collect proper image data set.

Future:

Explore other big companies API’s as Google or AWS, than some other providers as seeme.ai .

If you liked my post, have comments or recommendation of AI API’s, just give me info.

Welcome to a place where words matter. On Medium, smart voices and original ideas take center stage - with no ads in sight. Watch
Follow all the topics you care about, and we’ll deliver the best stories for you to your homepage and inbox. Explore
Get unlimited access to the best stories on Medium — and support writers while you’re at it. Just $5/month. Upgrade