Architecture of flat (top panel) and spectral (bottom panel) MultiPoSeIDoNs. The first one has a convolutional block, which produces eight feature maps. After that, the space dimensionality increases to 512 feature maps through five more convolutional blocks. The second one produces 9 and 576 feature maps in its first and last convolutional blocks, respectively. These layers are connected to deconvolutional ones, which decreases the space dimensionality to eight and nine feature maps in the last deconvolutional block for flat and spectral MultiPoSeIDoNs, respectively. Fine-grained features are added from each convolution to its corresponding deconvolution in both neural networks.