What helps U-Net to preserve information at low-level features?

U-Net preserves information on low-level features through the use of skip connections. These connections directly pass the outputs of the encoder layers (downsampling path) to the corresponding decoder layers (upsampling path), allowing the model to utilize both high-level and low-level features when reconstructing the image. This helps to avoid the loss of spatial information, which often occurs with the sequential application of convolutions and pooling.

It roughly looks like this: at each level of the encoder, the output is copied and concatenated with the input of the corresponding decoder level, which improves segmentation quality and detail of the result.