Mastering Machine Learning Algorithms

Atrous convolution

In some cases, a stride larger than one could be a good solution because it reduces the dimensionality and speeds up the training process, but it can lead to distorted images where the main features are not detectable anymore. An alternative approach is provided by the atrous convolution (also known as dilated convolution). In this case, the kernel is applied to a larger image patch, but skips some pixels inside the area itself (that's why someone called it convolution with holes). In the following graph, there's an example with (3 × 3) and dilation rate set to 2:

Example of atrous convolution with a Laplacian kernel

Every ...