Can the dimension of a feature map in the CNN be bigger than the original image dimension?

Question

I have implemented an convolutional Autoencoder of a paper. The thing is that I want to look at images of the size 32x32, but the original Autoencoder is used on 128x128 images. After I changed the Layers so it would fit my image size it worked fine with reconstruction and everything. But I now noticed that the dimension of several feature maps in the conv-layers are bigger than my original input images size.

For me it is unclear if a CNN can have a higher dimensional feature map than the input image. One of my conv-layers creates a feature map of the dimension 16x16x64 which is significantly higher than the input image size of 32x32x3. Can this happen in CNNs or is this some kind of no go?

Answer 1

The last argument is the number of feature maps produced, i.e. you have 64 feature maps of size 16x16. Each feature map is still smaller than your input image. The number of feature maps produced depends on the number of filters set in the convolutional layer.