How does filter size affect image understanding? For example, if our data is images of human faces, what is the difference between using a 3 * 3 filter and a 7 * 7 filter? Does increasing the size of the filter differentiate more shapes and textures?

CodePudding user response：

Yes, you are correct, to some extent.

Increasing the filter kernel allows the model to differentiate between more complex & bigger shapes and textures. For example:
Conv2D(filters=x, kernel_size=(7,7)) can differentiate more shapes than Conv2D(filters=x, kernel_size=(3,3)) can.