Can we do better than the basic, one-hidden layer autoencoder? We saw in the previous chapter, Chapter 4, Cats Versus Dogs – Image Classification Using CNNs, that deep CNNs perform well for image classification tasks. Naturally, we can apply the same concept for autoencoders too. Instead of using only one hidden layer, we use multiple layers (that is, a deep network) and instead of a fully connected dense layer, we use convolutional layers.
The following diagram illustrates the architecture of a deep convolutional autoencoder:
Constructing a deep convolutional autoencoder in Keras is simple. Once again, ...
