What is the primary motivation for using convolutional layers instead of fully connected layers for image data?