CRNet: A Detail-Preserving Network for Unified Image Restoration and Enhancement Task

TL;DR

CRNet tackles unified restoration and enhancement from multi-exposure inputs to produce HDR-like images by aligning five exposures and jointly refining high-frequency details. It introduces frequency separation via pooling and a Multi-Branch Block to fuse high- and low-frequency content, plus a Convolutional Enhancement Block with 7$\times$7 depthwise separable convolutions and an inverted bottleneck ConvFFN to enlarge receptive fields. The architecture processes concatenated, aligned frames and fuses refined features with a reference frame through three High-Frequency Enhancement Modules, guided by a mu-law tone-mapped L1 loss $L = \lVert T(H) - T(\hat{H}) \rVert_1$ with $T(x) = \frac{\log(1 + \mu x)}{\log(1+\mu)}$ and $\mu = 5000$. Evaluated on track 1 of the Bracketing Image Restoration Challenge, CRNet achieves third place and outperforms previous SOTA models in both PSNR_$\mu$ and SSIM_$\mu$, while maintaining lower computational costs. The results demonstrate improved edge and texture preservation, especially in high-frequency regions, enabling robust unified restoration of motion-blurred, noisy, and under/overexposed scenes.

Abstract

In real-world scenarios, images captured often suffer from blurring, noise, and other forms of image degradation, and due to sensor limitations, people usually can only obtain low dynamic range images. To achieve high-quality images, researchers have attempted various image restoration and enhancement operations on photographs, including denoising, deblurring, and high dynamic range imaging. However, merely performing a single type of image enhancement still cannot yield satisfactory images. In this paper, to deal with the challenge above, we propose the Composite Refinement Network (CRNet) to address this issue using multiple exposure images. By fully integrating information-rich multiple exposure inputs, CRNet can perform unified image restoration and enhancement. To improve the quality of image details, CRNet explicitly separates and strengthens high and low-frequency information through pooling layers, using specially designed Multi-Branch Blocks for effective fusion of these frequencies. To increase the receptive field and fully integrate input features, CRNet employs the High-Frequency Enhancement Module, which includes large kernel convolutions and an inverted bottleneck ConvFFN. Our model secured third place in the first track of the Bracketing Image Restoration and Enhancement Challenge, surpassing previous SOTA models in both testing metrics and visual quality.

Figures (9)

• Figure 1: Our CRNet can finely restore the details lost in images from multiple exposure inputs and enhance them into HDR image.
• Figure 2: An overview of the CRNet model reveals that it mainly comprises three components: alignment, high-frequency enhancement, and output fusion. In the High-Frequency Enhancement Module, we swiftly separate high and low-frequency features and then utilize cleverly designed Multi-Branch Block to fuse them. Subsequently, we employ a purely Convolutional Enhancement Block to efficiently extract and fuse features for image enhancement.
• Figure 3: We efficiently separate high and low-frequency information through simple pooling layers.
• Figure 4: Through asymmetric parallel convolutional groups, the model effectively integrates high and low-frequency information.
• Figure 5: Our Convolutional Enhancement Block utilizes depth-wise separable convolutions with large kernels to achieve a large receptive field. Additionally, it leverages ConvFFN with an inverted bottleneck structure to enhance feature fusion capability.