Non-Local Recurrent Network for Image Restoration (1806.02919v2)

Abstract: Many classic methods have shown non-local self-similarity in natural images to be an effective prior for image restoration. However, it remains unclear and challenging to make use of this intrinsic property via deep networks. In this paper, we propose a non-local recurrent network (NLRN) as the first attempt to incorporate non-local operations into a recurrent neural network (RNN) for image restoration. The main contributions of this work are: (1) Unlike existing methods that measure self-similarity in an isolated manner, the proposed non-local module can be flexibly integrated into existing deep networks for end-to-end training to capture deep feature correlation between each location and its neighborhood. (2) We fully employ the RNN structure for its parameter efficiency and allow deep feature correlation to be propagated along adjacent recurrent states. This new design boosts robustness against inaccurate correlation estimation due to severely degraded images. (3) We show that it is essential to maintain a confined neighborhood for computing deep feature correlation given degraded images. This is in contrast to existing practice that deploys the whole image. Extensive experiments on both image denoising and super-resolution tasks are conducted. Thanks to the recurrent non-local operations and correlation propagation, the proposed NLRN achieves superior results to state-of-the-art methods with much fewer parameters.

• The paper integrates non-local operations within an RNN to capture deep feature correlations for effective image restoration.
• The paper deploys an efficient RNN structure that propagates feature correlations across recurrent states, enhancing robustness in degraded conditions.
• The paper confines the non-local neighborhood to improve correlation accuracy, yielding superior restoration performance with fewer parameters.

An Overview of "Non-Local Recurrent Network for Image Restoration"

The paper "Non-Local Recurrent Network for Image Restoration" presents an innovative approach to image restoration by integrating non-local operations within a recurrent neural network (RNN) framework. This method addresses the challenge of capturing non-local self-similarity in images—a property that traditional deep networks often overlook.

Key Contributions

The proposed non-local recurrent network (NLRN) offers several notable contributions to the field of image restoration:

1. Integration of Non-Local Operations: Unlike traditional methods that isolate self-similarity measurement, the NLRN integrates a non-local module within deep networks. This allows for end-to-end training, effectively capturing deep feature correlation between pixels and their surrounding neighborhood.
2. Efficient RNN Structure: By utilizing the parameter-efficient RNN design, the NLRN propagates deep feature correlation across recurrent states. This approach enhances robustness against poor correlation estimates in degraded image conditions.
3. Confined Neighborhood for Correlation: The paper emphasizes the importance of limiting the neighborhood size for non-local operations, contrary to existing practices that use entire images. This confined approach improves the accuracy and reliability of correlation computations in degraded images.

Experimental Results

The authors provide extensive experimentation across image denoising and super-resolution tasks. The results demonstrate that the NLRN achieves superior restoration performance over state-of-the-art methods with significantly fewer parameters. Noteworthy improvements in PSNR and SSIM metrics highlight the effectiveness of the proposed network, particularly in challenging datasets like Urban100.

Implications and Future Directions

The integration of non-local operations in RNN frameworks opens new pathways for image restoration, merging the strength of self-similarity exploitation with deep learning. The proposed NLRN illustrates that deep networks can benefit from non-local image characteristics through strategic network design.

For future developments, this research suggests exciting possibilities:

• Broader Application Spectrum: The NLRN framework could be adapted to other domains where non-local characteristics are crucial, such as texture synthesis or medical imaging.
• Enhanced Network Architectures: Further exploration into network architectures with advanced forms of non-local operations may yield even greater gains in restoration quality.
• Scalability and Multimodal Extensions: The scalability of NLRN to handle larger datasets or different types of image degradation remains an attractive area for exploration. Extensions to multimodal data could also harness the power of non-local operations in more complex environments.

In summary, this research provides a robust framework for enhancing image restoration techniques by leveraging non-local operations within an RNN setting. The innovative approach and solid empirical results pave the way for further advancements in the integration of non-local image attributes with deep learning architectures.