The last operation we will introduce in this chapter is a bit different from the previous ones, as it is not meant to upsample a feature map provided. Instead, it was proposed to artificially increase the receptive field of convolutions without further sacrificing the spatial dimensionality of the data. To achieve this, dilation is applied here too (refer to the Transposed convolutions (deconvolution) section), though quite differently.
Indeed, dilated convolutions are similar to standard convolutions, with an additional hyperparameter, d, defining the dilation applied to their kernels. Figure 6-7 illustrates how this process does artificially increase the layer's receptive field:
Figure 6-7: Operations performed by a dilated-convolutional layer (defined here by a 2 × 2 kernel w, padding p = 1, stride s = 1, and dilation d = 2)
These layers are also called atrous convolutions, from the French expression à trous...
