In the decoder, the features from the unet blocks get concatenated with features from the encoder layer through 'skip connections'. The paper discusses how resc

In the decoder, the features from the unet blocks get concatenated with features from the encoder layer through 'skip connections'. The paper discusses how rescaling the backbone features (element-wise multiplication by some scalar) before concatenation improves image quality.