I am currently working on a semantic segmentation project based on satellite images,
I was doing some research and I chance upon this thing called gaussian filter where the weights of the
filter are concentrated in the middle. However I am confused as to what it really is and where do I apply this in my model. Is a gaussian filter/kernel a data augmentation? Is it the same as gaussian blur or gaussian noise? or issit something I have to imlpement on my conv layers when they have kernels inside. Hope someone can explain this to me thank you!
A Gaussian filter in image processing is also called Gaussian blur and is a low-pass filter.
You can apply it on your images to blur them, if you think it might be beneficial for the training.
It’s not the same as Gaussian noise, which is usually additive noise to the input signal.
You can also implement the Gaussian filters in a conv layer, but depending on your use case you might want to use
PIL methods to add the blur to the images.
Where did you read about these filters and did the source explain where and how these filters are applied?
I can’t really recall from where I read it, but yea was a little confused when doing my research on this. So like you mentioned, some of them were applying it as a data augmentation while some are applying it in their model when calling nn.Conv2D(…dilation=2). But I sort of understand the difference between these two already. Thank you for the explanation!
The case I am working with is a sen12ms dataset, which has 13 bands in a image. So I have to do my data augmenting on tensors instead of PIL image. As PIL can only work on a 3 channel image.