How to handle normalizing image with blank pixels in a Unet for image segmentation?

[Kevin_Sanchez]

Hello,
I am doing my first unguided project and just need a little advice. I’ve done a lot of reading and some online courses, but have not been able to find information on how to handle blank space. I do not want to tell exactly what I am doing, but here is some information that I can give:

-I have hundreds of images that are inconsistent shapes (curved edges, different sizes) and cropping part of the image is not an option. I put these images in a rectangular matrix so all my data is the same shape, but because of that I have a bunch of blank pixels.

-There is only 1 feature/class I am trying to model, but there is a large class imbalance in terms of the number of pixels(<1:100), so I am using a binaryfocalloss function. The range of values in the actual images is about 100-400. So I could set the blank pixels to zero then normalize or 100 then normalize. I am not sure it matters.

-For the training data, the features are determined by taking the difference in the values of 2 of the image channels and manually identifying the pattern of that feature.

Bonus question: If my channels have different value ranges (like 100-150 vs 300-400) should I normalize each channel by the range of that channel or the range of all the channels? I think the latter?

Thanks for your help!

[Mark_Daoust]

@Kevin_Sanchez

If my channels have different value ranges (like 100-150 vs 300-400) should I normalize each channel by the range of that channel or the range of all the channels? I think the latter?

If the ranges are “similar” it porobably doesn’t matter. But if the difference is large, I’d normalize by the channel’s individual range.

@Kevin_Sanchez

The range of values in the actual images is about 100-400. So I could set the blank pixels to zero then normalize or 100 then normalize. I am not sure it matters.

What?

@Kevin_Sanchez

how to handle blank space

@Kevin_Sanchez

There is only 1 feature/class I am trying to model

It sounds like you just need a binary image for the labels, and a logit-image for the predictions.

[Kevin_Sanchez]

Thank you for your reply.

@markdaoust

@Kevin_Sanchez

The range of values in the actual images is about 100-400. So I could set the blank pixels to zero then normalize or 100 then normalize. I am not sure it matters.

What?

To rephrase, what value should I give the blank pixels? Does it matter?

[Mark_Daoust]

@Kevin_Sanchez

To rephrase, what value should I give the blank pixels? Does it matter?

So the “on” pixels have a range of 100-400, okay, I think I’m with you now.

zero then normalize or 100 then normalize

One advantage to using 0 instead of 100, is that it’s easy to tell that zero is not in that valid range.
If you set the background pixels to 100, when you see a 100, it’s never clear if that’s a valid 100 or a background 100.

Another option here would be to add a valid [0,1] channel to the images so your network can see which pixels are valid, then the fill value would matter less.

[Kevin_Sanchez]

@markdaoust

So the “on” pixels have a range of 100-400, okay, I think I’m with you now.

Yes! sorry for the confusion.

@markdaoust

One advantage to using 0 instead of 100, is that it’s easy to tell that zero is not in that valid range.
If you set the background pixels to 100, when you see a 100, it’s never clear if that’s a valid 100 or a background 100.

Another option here would be to add a valid [0,1] channel to the images so your network, then the fill value would matter less.

Both of those ideas make sense. Thank you for your help!!

[Umar_Sadique]

Hi Dear!
I read you answer.
I know you can help me to solve my problem.
I predicted an image using Unet the image show blank in yellow colour. I multiply 255 to change pixel values. Now the pixel values are from 180 to 255 with single channel.
Now I want to convert the image into black and white to draw contour on original image.

How I can convert the image into black and white(mask).