Abstract
This paper proposes Deep Bi-Dense Networks (DBDN) for single imagesuper-resolution. Our approach extends previous intra-block dense connectionapproaches by including novel inter-block dense connections. In this way,feature information propagates from a single dense block to all subsequentblocks, instead of to a single successor. To build a DBDN, we firstly constructintra-dense blocks, which extract and compress abundant local features viadensely connected convolutional layers and compression layers for furtherfeature learning. Then, we use an inter-block dense net to connect intra-denseblocks, which allow each intra-dense block propagates its own local features toall successors. Additionally, our bi-dense construction connects each block tothe output, alleviating the vanishing gradient problems in training. Theevaluation of our proposed method on five benchmark datasets shows that ourDBDN outperforms the state of the art in SISR with a moderate number of networkparameters.