RDB-FSU: Residual Dense Network with Feature Selection Unit for Image Super Resolution

Abstract

After the great success of deep convolution neural network (DCNN). In practical, residual and dense networks. Most deep methods do not make full use of features from the original low-resolution (LR) images. This paper proposed a novel residual in dense network with feature selection unit (RDB-FSU) that tackle this problem in both image and video SR. RDB-FSU fully use the features from all the convolution layers. Specifically, the residual dense block (RDB) presented to extract features via dense connected convolution layers. Then feature selection unit (FSU) is adaptively learn more effective features from extracted features to be up scaled using pixel-shuffle module. Extensive experiments with various models show that the presented RDB-FSU fulfills appropriate performance versus state-of-the-art techniques on both image and video benchmark dataset.