Fig. 4.

Illustration of Y-Net architecture. Resizing operations are performed in the contracting path which is responsible for learning the feature representation in measured data. Then, the feature maps in each spatial level of the two contracting paths are concatenated to their respective spatial levels in the decoder. Furthermore, an auxiliary loss is included at the bottom of the contracting path responsible for learning the texture information. The spatial dimension of feature maps is indicated above the corresponding cube.