Laplacian-Mamba: Mamba-Based Laplacian Pyramid Enhancement Network for Unpaired High-Definition Images
Zhiquan Mao, Mengning Yang · 2025
The domain of image enhancement for high-definition images has attracted widespread attention, as they become increasingly prevalent. However, the increased pixel count of high-definition images brings a huge burden to computation and memory. To tackle the issue, most of the existing methods employ high-magnification downsampling, which inevitably results in information loss. Excessive information loss will lead to artifacts, which significantly impacts the visual quality of the enhanced results. In addition, the reliance on paired data remains a major challenge in the field of image enhancement. In this work, we focus on designing a method that can both achieve realistic enhancement and eliminate data dependency based on closed-form Laplacian pyramid decomposition and reconstruction. To this end, we propose Laplacian-Mamba, a Mamba-based Laplacian pyramid enhancement network for unpaired high-definition images. Specifically, we proposed a High-Frequency Preservation Module (HF-PModule) by improving SSMs to focus on restoring the information of high-frequency sub-bands. Moreover, to enhance the brightness of the images towards the target domain and preserve global information of low-frequency component, we proposed a Low-Frequency Enhancement Module (LFEModule) which integrates a histogram transformer block and four Mamba layers. Through comprehensive evaluation, our method has demonstrated excellent performance.