Depending on the approach, the increase of the receptive field might reduce the effect of fine-grained details. E.g. when increasing the stride or adding more pooling layers, fine-grained details can be lost due to downsampling.
- Deeper Network
- Larger kernels - more params
- Strides greater than 1 - downsampling
- Dilated Convolutions - increasing the receptive field w/o downsampling
- ...
- Tanh regularization of scaling
- Soft training in CNN by adding a context channel