Residual Dense Network for Image Restoration

Convolutional neural network has recently achieved great success for imagerestoration (IR) and also offered hierarchical features. However, most deep CNNbased IR models do not make full use of the hierarchical features from theoriginal low-quality images, thereby achieving relatively-low performance. Inthis paper, we propose a novel residual dense network (RDN) to address thisproblem in IR. We fully exploit the hierarchical features from all theconvolutional layers. Specifically, we propose residual dense block (RDB) toextract abundant local features via densely connected convolutional layers. RDBfurther allows direct connections from the state of preceding RDB to all thelayers of current RDB, leading to a contiguous memory mechanism. To adaptivelylearn more effective features from preceding and current local features andstabilize the training of wider network, we proposed local feature fusion inRDB. After fully obtaining dense local features, we use global feature fusionto jointly and adaptively learn global hierarchical features in a holistic way.We demonstrate the effectiveness of RDN with several representative IRapplications, single image super-resolution, Gaussian image denoising, imagecompression artifact reduction, and image deblurring. Experiments on benchmarkand real-world datasets show that our RDN achieves favorable performanceagainst state-of-the-art methods for each IR task quantitatively and visually.