Residual Dense Network for Image Super-Resolution

Very deep convolutional neural networks have recently achieved great success for image super-resolution and offer hierarchical features, but most deep CNN-based SR models do not make full use of the hierarchical features from the original low-resolution images, achieving relatively low performance. The paper proposes a Residual Dense Network (RDN) that fully exploits the hierarchical features from all convolutional layers, using a residual dense block (RDB) to extract abundant local features via densely connected convolutional layers and direct connections from preceding RDBs to all layers of the current RDB, forming a contiguous memory mechanism, with local feature fusion adaptively learning effective features and stabilizing training of wider networks.

After fully obtaining dense local features, RDN uses global feature fusion to jointly and adaptively learn global hierarchical features in a holistic way. Experiments on benchmark datasets with different degradation models show that RDN achieves favorable performance against state-of-the-art methods, demonstrating the benefit of combining local and global hierarchical feature reuse for super-resolution.