Queries for Custom Vision API

Question

Hi, I would be using the Custom Vision API using Python SDK for my commercial application purposes and would love few suggestions :

1- I'd be using python for development, and my data resides in say, azure blob storage. Will i be able to get a jupyter notebook inside azure environment

2- I'd be using the custom model, and before that I will have to preprocess the images, are there any preprocessing capibilities also provided by vision AI?

3- Can the model be retrained after new images are incorporated? I just saw a question from Nov 2023 saying custom vision was not allowing retraining

4- Lets say instead of using service via python sdk, i decide to use custom vision studio, will i be able to do image preprocessing on the studio itself, and then export the results as well for post processing/ do post processing on studio itself?

5- I have to analyze the images i upload to find if magazine X(only X, maybe different versions of X but no other magazine), so i have only 1 no of classes for the detection. So will the model be able to work with this ? I mean i could annotate the regions of my X magazine in the training images, do i have to annotate everything else as Other?

@navba-MSFT or any other azure expert, please help at the earliest!
Thanks a lot!

Answer 1

@Rishi Khemka

Thanks for reaching out to us, for a quick answer of your questions.

1. Yes, you can use Jupyter notebooks within Azure. Azure provides a service called Azure Notebooks, which is a free hosted Jupyter Notebook service that you can use for simple cloud-based data exploration. Doc - https://learn.microsoft.com/en-us/azure/machine-learning/how-to-run-jupyter-notebooks?view=azureml-api-2
2. Azure's Custom Vision service does not have specific image preprocessing capabilities built-in. You would typically handle image preprocessing (resizing, normalization, augmentation etc.) in your own application before you send the images to the Custom Vision service.
3. Yes, you can retrain your models in the Custom Vision service. This allows you to make improvements over time as you get more labeled images to train with. Note that retraining a model will replace the previous version of the model. Doc - https://learn.microsoft.com/en-us/azure/ai-services/custom-vision-service/test-your-model
4. The Custom Vision Studio provides a user-friendly interface for building and training your custom vision models, but it does not have built-in image preprocessing capabilities. You can export the results, and any post-processing would need to be done in your own application or a separate tool.
5. The Custom Vision service is designed to handle multi-class classification, but it can also work with a single class. If you're only interested in identifying whether an image contains a specific object (magazine X in your case), you can simply train a model with images of that object. You do not necessarily need to annotate everything else as "Other", unless you want your model to explicitly identify images that do not contain the object of interest. In your case, you could annotate regions of the magazine X in the training images and the model will learn to recognize those features. Images that do not contain those features would not be classified as magazine X.

Please consider your scenario with more details and situation evaluation. Let us know if you have more questions.

regards,

Yutong