DRWKV: Focusing on Object Edges for Low-Light Image Enhancement

TL;DR

The paper tackles edge fidelity in low-light image enhancement by introducing DRWKV, a framework that decouples illumination from edge structures via Global Edge Retinex theory. It introduces Evolving Scanning RWKV to better model edge continuity, and Bi-SAB with MS2-Loss to coherently align brightness and color while suppressing artifacts. Across five benchmarks, DRWKV delivers leading PSNR, SSIM, and NIQE with low computational cost and extends its benefits to low-light tracking tasks, validating generalization. The work also provides extensive architecture details and supplementary experiments to support robustness and potential mobile deployment.

Abstract

Low-light image enhancement remains a challenging task, particularly in preserving object edge continuity and fine structural details under extreme illumination degradation. In this paper, we propose a novel model, DRWKV (Detailed Receptance Weighted Key Value), which integrates our proposed Global Edge Retinex (GER) theory, enabling effective decoupling of illumination and edge structures for enhanced edge fidelity. Secondly, we introduce Evolving WKV Attention, a spiral-scanning mechanism that captures spatial edge continuity and models irregular structures more effectively. Thirdly, we design the Bilateral Spectrum Aligner (Bi-SAB) and a tailored MS2-Loss to jointly align luminance and chrominance features, improving visual naturalness and mitigating artifacts. Extensive experiments on five LLIE benchmarks demonstrate that DRWKV achieves leading performance in PSNR, SSIM, and NIQE while maintaining low computational complexity. Furthermore, DRWKV enhances downstream performance in low-light multi-object tracking tasks, validating its generalization capabilities.

Figures (9)

• Figure 1: Comparison of our DRWKV with different methods: (a) DRWKV achieves the best PSNR in benchmark tests; (b) DRWKV balances efficiency, parameters, and computational load; (c) DRWKV exhibits more robust performance in edge details.
• Figure 2: The overview of the DRWKV model. (a) DRWKV processes low-light images through two steps: (i) Light Preprocessing and (ii) Deep Detail Mining. The positions of Block1 and Block2 correspond to the layout settings of the two Retinex theories. (b) The block structure of Evolving Scanning RWKV, which embeds two core mixing modules: Evolving Scanning Spatial Mix and Channel Mix. (c) The block structure of the Bilateral Spectrum Aligner Block.
• Figure 3: Visualization results of RetinexNet, RRDNet, RetinexMamba, RetinexFormer, and DRWKV on the LOLv2-Real, LOLv2-Synthetic, and LSRW-Huawei datasets respectively (zoom in to display details for each model).
• Figure 4: An ablation study is conducted on the proposed decompositional consistency loss $L_{recon}$, edge sparsity loss $L_{sparse}$, illumination smoothness loss $L_{smooth}$, artifact suppression loss $L_{artifact}$, and parameter regularization loss $L_{reg}$.
• Figure 5: Evolving Scanning RWKV's structure, it consists of two parts: Evolving Scanning Spatial Mix and Channel Mix.