Would the Object Detection model work here?

[PVosEska]

Hi Guys,

I am currently investigating the AI builder models en trying to find use-cases for this. 

I currently work at a cardboard manufacturer and one of the struggles they have is classifying 'cloudiness' of cardboard. This is to say the flatness or evenness of the cardboard. 

Pictures are taken from samples with a camera straight from the top, with a flash. the amount of light-'particles' visible will determine the classification is. The more visible light particles, the worse the grade of the paper, as more reflections show there is more unevenness.

My question:

Can you teach the Object detection model to recognize tens, if not hundreds of of the same objects (light particles in this case) in one picture or will this not be viable with these models?

Hope to hear what you think!

Regards,

Peter

[cohenjacobd · Accepted solution]

Hi Peter! 

Jake here, a PM on the AI Builder team. Are you looking for the model to provide the # of light particles in a given image of cardboard? Or to just classify if an image of cardboard is flat vs not, or on a scale of flat, somewhat flat, very much not flat? 

If it's the latter, this sounds like a great fit for Image Classification. Instead of labeling boxes around each light reflection, you can label images as "flat" "somewhat flat" and "very much not flat" given the # of light particles in the entire image. This way the model starts to learn that the amount of light reflection is what classifies an image as flat vs somewhat flat vs very much not flat. 

I would recommend using Lobe (https://lobe.ai), a free easy-to-use desktop app by Microsoft to train custom machine learning models. If you can download these images of cardboard to your desktop, you can easily import and label the images and it will automatically train a model. I would recommend starting with a hundred images in those 3 groups to get a proof of concept and see how it performs.

If you'd like to chat more about your business scenario or if you have more questions about Lobe, feel free to email us aihelpen@microsoft.com.

[PVosEska]

Hi @cohenjacobd ,

Thank you for your reply. I thought about it a bit more, and i thought of a better analogy of what we're trying to accomplish.

Imagine snow on the tv-screen. Black and white particles, imagine that the screen would be fully black without the snow and the snow particles are white. 

This directly translates to my use-case as the images to be measured are also black and white. If there is no snow in the image, this indicates the surface is smooth and the quality is perfect. 

Every snowflake that is on the image should be counted. This way, we can create a ranking similar to":

#1: 0 - 10 snowflakes,

#2: 10 - 50 snowflakes

#3: 50 - 100 snowflakes

and so on...

I would add some images here, but have yet to receive one sadly.

I am going to try import some images in Lobe as soon as i receive them and see if it will help us! Thanks for taking the time to respond, much appreciated!

Regards,

[cohenjacobd]

Thanks for sharing more details. Is the intent for the model to categorize "good" vs "bad" cardboard? Like the good cardboard goes off the manufacturing line for sale, and the bad cardboard is recycled or used for something else, or sold for less? Or is this to alert the manufacturing line for anomalies when the occur? Or just overall analysis to assess how well the manufacturing of cardboard is going over time?

Please do let us know how it goes training a model with Lobe. If having 3 categories based of low, medium, high amount of snowflakes is sufficient for your needs, then I think image classification is a good starting point. 

Jake

[PVosEska]

Hi @cohenjacobd ,

The intent is for our R&D department when developing new types of board and to increase quality of existing board. It's a way of objectifying the roughness of the board. Up untill now it is only measured in the eye of the beholder. 

I was expecting to be able to import the Lobe model in the Power Platform AI model, but it doesn't seem this is a possibility? This means that i would have to build my own application around it. I have some developement skills, but have to see if i can tackle this.

I did some testing and it seems the model works really well! As i haven't received any testmaterial yet, I did some experiments on my own, and it seems it correctly predicts what grade it belongs in, 97% of the time (Grade 1 - 4 in my case, where 1 is bad and 4 is really well).

Next step is actual testing, have to wait for that until next week.

Thanks for your help so far!

Regards,

[cohenjacobd]

Great to hear your testing so far is working well! Excited to hear how the test material works next week. 

We are actively developing the support to import a Lobe model to Power Platform, and will start a private preview soon 🙂 Would you like to join the private preview? Our team can provide assistance in building your business solution. 

Jake

[PVosEska]

I would love to join the private preview! 

I have a couple of years of experience with building apps and flows so that should not be an issue. 

Let me know if you need anything for me to join the preview.

Thank you for the invitation!

Regards,

Peter

[cohenjacobd]

Hi Peter! I have sent you a personal message inviting you to the private preview. Please check that out and sign up and we can follow up more. I'm looking forward to hearing more about your progress 🙂 

Jake