Since the hidden layers of a CNN work as a trainable feature extractor, more detailed content based on a larger number of pixels shall require bigger filter sizes. But for cases where localized differences are to receive greater attention, smaller filter sizes are required.

I know there is a lot of topic on the internet regarding CNN and most of them have a simple explanation about Convolution Layer and what it is designed for, but they don’t explain

How many convolution layers are required?

What filters should I use in those convolution layers?

  • $\begingroup$ There's this question about how to choose the size of the kernel. This is related to your second question, or is it the same question? The filter is only specified by its size, so I guess you're asking the same question as in the linked post. So, maybe you should also remove that question and just ask "how many layers". $\endgroup$
    – nbro
    May 19, 2020 at 20:06


You must log in to answer this question.