Residual Dense Network for Image Super-Resolution

[harshraj22]

RDN 📓

CVPR 2018

Architectural Details

• They propose skip connections, arguing it to preserve local and global features extracted by the CNN filters, being necessary for obtaining good quality super resoluted image.

• Each filter in the Residual Dense Block has skip connections from all previous filters, where the outputs are concatenated before being passed on to the filter.

• While local features are passed on to the next layers through skip connections in Residual Dense Block, the global features are passed through the Longer skip connection. The authors argue that this helps in using the Hierarchical features extracted by the initial filters, which plays an important role in generating the Super Resolution Image.

Dataset Details:

• The model was trained on 800 images of DIV2K dataset. SET5 dataset was used for validation.
• They randomly extract 16 LR RGB patches with the size of 32 × 32 as inputs and randomly augment the patches by flipping horizontally or vertically and rotating 90 degrees.
• The authors train the model for 200 epochs which takes 1 day on a Titan Xp GPU.