Abstract:Most existing deep learning-based image restoration methods usually aim to remove degradation with uniform spatial distribution and constant intensity, making insufficient use of degradation prior knowledge. Here we bootstrap the deep neural networks to suppress complex image degradation whose intensity is spatially variable, through utilizing prior knowledge from degraded images. Specifically, we propose an ingenious and efficient multi-frame image restoration network (DparNet) with wide & deep architecture, which integrates degraded images and prior knowledge of degradation to reconstruct images with ideal clarity and stability. The degradation prior is directly learned from degraded images in form of key degradation parameter matrix, with no requirement of any off-site knowledge. The wide & deep architecture in DparNet enables the learned parameters to directly modulate the final restoring results, boosting spatial & intensity adaptive image restoration. We demonstrate the proposed method on two representative image restoration applications: image denoising and suppression of atmospheric turbulence effects in images. Two large datasets, containing 109,536 and 49,744 images respectively, were constructed to support our experiments. The experimental results show that our DparNet significantly outperform SoTA methods in restoration performance and network efficiency. More importantly, by utilizing the learned degradation parameters via wide & deep learning, we can improve the PSNR of image restoration by 0.6~1.1 dB with less than 2% increasing in model parameter numbers and computational complexity. Our work suggests that degraded images may hide key information of the degradation process, which can be utilized to boost spatial & intensity adaptive image restoration.

Adaptive Residual Networks for High-Quality Image Restoration

Hypernetwork-Based Adaptive Image Restoration

Residual Dense Network for Image Restoration

Revitalizing Convolutional Network for Image Restoration

Deep Likelihood Network for Image Restoration with Multiple Degradation Levels

A Comprehensive Review of Deep Learning-Based Real-World Image Restoration

Convolutional Neural Network Combined with Half-Quadratic Splitting Method for Image Restoration

Learning Deformable and Attentive Network for image restoration

Image Restoration Using Convolutional Auto-encoders with Symmetric Skip Connections.

A survey of deep learning approaches to image restoration

Two-stream deep sparse network for accurate and efficient image restoration

A Double Feature Fusion Network with Progressive Learning for Sharper Inpainting

Densely Connected High Order Residual Network for Single Frame Image Super Resolution

Dilated Residual Encode-Decode Networks for Image Denoising

Image Restoration Using Deep Regulated Convolutional Networks

Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Symmetric Skip Connections.

Attention-Based Real Image Restoration

Image Restoration Via Deep Memory-Based Latent Attention Network

Unified-Width Adaptive Dynamic Network for All-In-One Image Restoration

Denoising Prior Driven Deep Neural Network for Image Restoration

Wide & deep learning for spatial & intensity adaptive image restoration