Image Restoration Network with Adaptive Channel Attention Modules for Combined Distortions

Haeyun Lee, Sunghyun ChoJournal of the Korea Computer Graphics Society 2019 · DOI

Abstract

Images obtained from systems such as autonomous driving cars or fire-fighting robots often suffer from several degradations, including noise, motion blur, and compression artifacts. It is difficult to apply image recognition to these degraded images, making image restoration essential. However, these systems often cannot recognize the specific type of degradation, complicating the restoration process. In this paper, we propose a deep neural network that restores natural images from multiple unknown distortions. By adopting channel attention modules and skip connections, the network focuses on information vital to restoration. Experimental results show that the proposed method is simpler to train and outperforms existing state-of-the-art methods.

Introduction

In controlled laboratory settings, image restoration models typically specialize in a single task, such as denoising or deblurring. However, real-world applications like autonomous vehicles encounter "combined distortions" where multiple factors degrade an image simultaneously. The primary challenge in these scenarios is that the system rarely knows the type or severity of the distortion beforehand.

Existing solutions often require manual parameter input or utilize computationally expensive architectures, such as reinforcement learning, to estimate distortion types. This paper introduces an end-to-end residual learning architecture that automatically adapts to various distortions using an adaptive channel attention mechanism.

Proposed Methodology

The proposed network follows a sequential architecture designed for efficiency and dynamic adaptation. It consists of three primary stages: feature extraction, channel attention blocks, and feature reconstruction.

Overall architecture of the proposed image restoration network, showing the sequential flow from input to the final restored result.

Adaptive Channel Attention

The core innovation lies in the Channel Attention (CA) modules embedded within five sequential blocks. Each block utilizes four convolutional layers followed by a CA module to prioritize feature channels based on the specific degradation present in the input.

Detailed structure of the Channel Attention (CA) module, illustrating the use of Global Average Pooling (GAP) and 1x1 convolutions to generate channel weights.

As shown in the detailed module structure, the process begins with Global Average Pooling (GAP), which compresses spatial information into a channel descriptor. This descriptor passes through two $1 \times 1$ convolutions with ReLU and Sigmoid activations to calculate a weight for each channel. Channels containing information critical for fixing a specific distortion (e.g., high-frequency details for blur) receive higher weights, while less useful channels are suppressed. This allows the network to dynamically specialize its processing on the fly.

Residual Learning

The network employs Residual Learning, meaning it is trained to predict the "noise" or distortion component rather than the clean image itself. The final output, $\tilde{x}$, is the sum of the original degraded input and the network's predicted correction, which facilitates faster convergence and more stable training.

Experiments and Results

Dataset and Training

The model was trained using the DIV2K dataset. To ensure robustness, the authors performed extensive data augmentation, including downsampling and cropping, resulting in over 230,000 training patches. Distortions were synthesized by applying varying levels of Gaussian blur, Gaussian noise, and JPEG compression in combinations ranging from "Extremely Mild" to "Extremely Severe."

Performance Comparison

The proposed method was evaluated against state-of-the-art models, including DnCNN, RL-Restore, and OWAN, using PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity) metrics.

  • Quantitative Results: In the "Mild" distortion category, the proposed model achieved a PSNR of 28.54 dB, outperforming OWAN (28.33 dB) and RL-Restore (28.04 dB). This performance advantage was maintained across moderate and severe distortion levels.
  • Qualitative Insights: Visual analysis indicates that the proposed network produces sharper edges and more natural textures compared to DnCNN (which tends to over-blur) and RL-Restore (which often leaves ringing artifacts).

Conclusion

By integrating adaptive channel attention modules into a residual learning framework, this research provides a powerful yet simple solution for blind image restoration. The model's ability to dynamically reweight features allows it to handle complex, combined distortions without prior knowledge of the degradation parameters. Future work will focus on network quantization and compression to enable real-time deployment on low-power edge devices, such as drones and mobile robots.

Summary updated

Abstract

The image obtained from systems such as autonomous driving cars or fire-fighting robots often suffer from several degradation such as noise, motion blur, and compression artifact due to multiple factor. It is difficult to apply image recognition to these degraded images, then the image restoration is essential. However, these systems cannot recognize what kind of degradation and thus there are difficulty restoring the images. In this paper, we propose the deep neural network, which restore natural images from images degraded in several ways such as noise, blur and JPEG compression in situations where the distortion applied to images is not recognized. We adopt the channel attention modules and skip connections in the proposed method, which makes the network focus on valuable information to image restoration. The proposed method is simpler to train than other methods, and experimental results show that the proposed method outperforms existing state-of-the-art methods.
BibTeX
@article{lee2019image,
  author = {Haeyun Lee and Sunghyun Cho},
  title = {{Image Restoration Network with Adaptive Channel Attention Modules for Combined Distortions}},
  journal = {Journal of the Korea Computer Graphics Society},
  year = {2019},
  doi = {10.15701/kcgs.2019.25.3.1},
}
Figure