Mastering Computer Vision with TensorFlow 2.x

Atrous convolution

We introduced the concept of convolution in Chapter 4, Deep Learning on Images, but we did not get into the various types of Atrous convolution. Atrous convolution, also known as dilated convolution, increases the field of view of convolution. The traditional CNN uses max pooling and stride to quickly reduce the size of a layer but in doing so, it also reduces the spatial resolution of the feature maps. Atrous convolution is a method that's used to get past this problem. It does this by modifying the stride using the Atrous value, thus effectively changing the filter's field of values, as shown in the following diagram:

The preceding diagram shows the Atrous convolution with rate = 2. It skips every other cell compared ...