Abstract
Generating high-quality high-dynamic-range (HDR) images in dynamic scenes remains a challenging task. Recently, Transformers have been introduced into HDR imaging and have demonstrated superior performance over traditional convolutional neural networks (CNNs) in handling large-scale motion. However, due to the low-pass filtering nature of self-attention, Transformers tend to weaken the capture of high-frequency information, which impairs the recovery of structural details. In addition, their high computational complexity limits practical applications. To address these issues, we propose HL-HDR, a high-low-frequency-aware ghost-free HDR reconstruction network for dynamic scenes. By decomposing features into high- and low-frequency components, HL-HDR effectively overcomes the limitations of existing Transformer and CNN-based methods. The Frequency Alignment Module (FAM) captures large-scale motion in the low-frequency branch while refining local details in the high-frequency branch. The Frequency Decomposition Processing Block (FDPB) fuses local high-frequency details and global low-frequency context, enabling precise HDR reconstruction. Extensive experiments on five public HDR datasets demonstrate that HL-HDR consistently outperforms state-of-the-art methods in both quantitative metrics and qualitative evaluation.