Residual Dense Network for Image Restoration

TL;DR

This paper introduces a residual dense network (RDN) that fully exploits hierarchical features from all convolutional layers to improve image restoration tasks such as super-resolution, denoising, and deblurring, achieving state-of-the-art results.

Contribution

The paper proposes a novel residual dense block (RDB) with local feature fusion and a global feature fusion mechanism to enhance feature utilization and training stability in deep CNNs for image restoration.

Findings

RDN outperforms existing methods on benchmark datasets.

RDN achieves superior visual quality in restored images.

The approach is effective across multiple image restoration tasks.

Abstract

Convolutional neural network has recently achieved great success for image restoration (IR) and also offered hierarchical features. However, most deep CNN based IR models do not make full use of the hierarchical features from the original low-quality images, thereby achieving relatively-low performance. In this paper, we propose a novel residual dense network (RDN) to address this problem in IR. We fully exploit the hierarchical features from all the convolutional layers. Specifically, we propose residual dense block (RDB) to extract abundant local features via densely connected convolutional layers. RDB further allows direct connections from the state of preceding RDB to all the layers of current RDB, leading to a contiguous memory mechanism. To adaptively learn more effective features from preceding and current local features and stabilize the training of wider network, we proposed local feature fusion in RDB. After fully obtaining dense local features, we use global feature fusion to jointly and adaptively learn global hierarchical features in a holistic way. We demonstrate the effectiveness of RDN with several representative IR applications, single image super-resolution, Gaussian image denoising, image compression artifact reduction, and image deblurring. Experiments on benchmark and real-world datasets show that our RDN achieves favorable performance against state-of-the-art methods for each IR task quantitatively and visually.