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.
