FreqMamba: exposure-aware frequency decomposition with interleaved state scanning for HDR imaging

Abstract

Multi-exposure High Dynamic Range (HDR) image reconstruction aims to merge multiple Low Dynamic Range (LDR) images into a single HDR image with extended luminance range. Existing methods typically model spatial relationships and cross-exposure complementarity in isolation, overlooking their intrinsic coupling—where information reliability at each spatial location is fundamentally determined by its exposure level. Meanwhile, binary exposure masks widely used for region selection introduce boundary artifacts incompatible with the continuous variation of scene radiance, and frequency components are processed uniformly without accounting for their exposure-dependent signal quality. To address these limitations in a unified manner, we propose FreqMamba, a tri-branch frequency-aware framework built upon the synergy of CNNs, Swin Transformers, and Mamba. At its core, Multi-Exposure Interleaved State Scanning (MEISS) interleaves tokens from different exposures within quad-directional Mamba scanning, enabling joint exposure-spatial modeling with linear complexity. Luminance-Adaptive State Gating (LASG) further replaces hard threshold masks with a learnable soft gating mechanism that modulates state transitions based on multi-exposure luminance statistics, ensuring smooth radiance continuity. Building on this, Exposure-aware Frequency Decomposition and Recoupling (EFDR) decomposes features into frequency sub-bands processed by architecturally matched branches and fuses them via physics-guided reliability estimation under radiance consistency constraints. Extensive experiments on the Kalantari and Tel benchmarks demonstrate that FreqMamba achieves 42.43 dB PSNR-l and 0.9883 SSIM-l on challenging dynamic scenes, outperforming state-of-the-art methods with only 0.97M parameters and O(3N) linear complexity for cross-exposure modeling.