PGHDR: Dynamic HDR reconstruction with progressive feature alignment and quality-guided fusion

High Dynamic Range (HDR) imaging captures the complete luminance information of a scene by fusing a multi-exposure image sequence. However, a key challenge remains: ghosting artifacts caused by object motion in dynamic scenes. Existing methods typically adopt an align-then-fuse strategy, often overlooking the spatial variability of alignment quality, which makes it difficult to balance ghosting suppression and detail preservation when handling complex motion. To address this issue, this paper proposes PGHDR, which leverages a quality-aware progressive alignment strategy. The core innovation is a learning paradigm that jointly optimizes motion estimation and alignment quality perception. This approach treats alignment quality as a learnable, spatially-varying confidence and materializes this concept through three synergistic modules. The Progressive Deformable Feature Alignment (PDFA) module achieves robust feature extraction through a two-stage deformable convolution and exposure-aware modulation. The Quality-Guided Motion Compensation (QGMC) module jointly learns optical flow and a confidence mask, enabling reliability-aware motion estimation. The Structure-Preserving Adaptive Fusion (SPAF) module performs weighted fusion based on the learned quality information while preserving image details via a local structure constraint. Experiments show that PGHDR achieves a PSNR- $$\ell $$ of 42.58 dB and an HDR-VDP-2 score of 67.15. Visual evaluations on multiple datasets demonstrate that our method significantly outperforms existing approaches, achieving high-quality, ghost-free HDR reconstruction.