Fig. 1.

The considered DDCAE reference architecture. The input noise is present only in the DDCAE training phase while it is zeroed in the validation and test phases. Furthermore, depending on the actually considered DDCAE architecture, the innermost flattening and dense layers may be absent. Accordingly, $L+1$ is the depth of the Encoder, while $L$ is the number of the corresponding Convolutional+Pooling+Batch Normalization (BN) layers. Finally, the taxonomy: $M\times N\times F$ indicates a convolutional layer (or a filter kernel or an input image) of spatial dimension: $M\times N$ which embraces $F$ feature maps.