Bracketing Image Restoration and Enhancement with High-Low Frequency Decomposition

TL;DR

HLNet tackles the challenge of real-world image degradations by unifying restoration and enhancement through a dual-module architecture that separates common and degradation-specific processing. It introduces a Spatial-Channel Enhancement Block (SCEB) to capture joint spatial-channel cues and a High-Low Frequency Decomposition Block (HLFDB) to tailor high- and low-frequency information to different degradations, processing five differently exposed frames in a sequence after feature alignment. The method employs gamma-corrected concatenation inputs, a sequential processing pipeline, and a tonemapped $L_1$ loss to train on HDR-domain data, with wavelet-based multi-scale fusion and transformer-based channel attention in the low-frequency path. Empirically, HLNet achieves state-of-the-art or competitive results on the BracketIRE+ Track 2 dataset, including a ~0.75 dB PSNR-μ improvement over a baseline and a top-tier NTIRE 2024 challenge performance, demonstrating improved detail recovery and structural fidelity across diverse degradations with practical computational efficiency.

Abstract

In real-world scenarios, due to a series of image degradations, obtaining high-quality, clear content photos is challenging. While significant progress has been made in synthesizing high-quality images, previous methods for image restoration and enhancement often overlooked the characteristics of different degradations. They applied the same structure to address various types of degradation, resulting in less-than-ideal restoration outcomes. Inspired by the notion that high/low frequency information is applicable to different degradations, we introduce HLNet, a Bracketing Image Restoration and Enhancement method based on high-low frequency decomposition. Specifically, we employ two modules for feature extraction: shared weight modules and non-shared weight modules. In the shared weight modules, we use SCConv to extract common features from different degradations. In the non-shared weight modules, we introduce the High-Low Frequency Decomposition Block (HLFDB), which employs different methods to handle high-low frequency information, enabling the model to address different degradations more effectively. Compared to other networks, our method takes into account the characteristics of different degradations, thus achieving higher-quality image restoration.

Figures (6)

• Figure 1: Our HLNet excels in restoring image details and texture, effectively enhancing the edge details of the stars in the figure.
• Figure 2: Overview of HLNet. In HLNet, feature alignment is performed first, followed by the gradual feeding of each frame into the network. The feature extraction stage consists of shared weight modules and non-shared weight modules. Each frame needs to pass through both modules, first through the shared weight module and then through the non-shared weight module.
• Figure 3: The architecture of the proposed High-Low Frequency Decomposition Block starts by using Average pooling to extract the low-frequency information from the features. Then, subtracting this low-frequency information from the overall features yields the high-frequency information.
• Figure 4: In the architecture of the proposed Global Feature Extraction Block, three successive downsampling operations are required. To prevent the loss of structural information during downsampling, the module employs cross-scale feature fusion based on wavelet transform.
• Figure 5: Examples of comparisions on the track 2 of the Bracketing Image Restoration and Enhancement Challenge dataset.