Digital camera sensors face a fundamental physical constraint: they cannot simultaneously capture detail in deep shadows and bright highlights. To overcome this, High Dynamic Range (HDR) imaging typically merges multiple Low Dynamic Range (LDR) frames taken at different exposure levels.
While this works perfectly for static scenes, it fails in the real world. If an object moves or the camera shakes between shots, the pixels no longer align. When a fusion algorithm tries to blend these frames, it creates "ghosting"—translucent, duplicated, or blurry artifacts where the object used to be. Recent research has shifted from simple pixel-blending to sophisticated "ghost-free" reconstruction techniques that treat HDR imaging as a joint problem of alignment, hallucination, and fusion.
The Spatial Temporal Mismatch
The core challenge of HDR imaging in dynamic scenes is that visual information varies drastically across exposures. In a short exposure, a moving car might be sharp but underexposed; in a long exposure, that same car is well-lit but motion-blurred.
Traditional alignment methods like optical flow often struggle here because the "features" they look for—edges, textures, and corners—change appearance or disappear entirely between exposures. As noted in Optical flow in the presence of spatially-varying motion blur by Travis Portz, trying to map sharp features to blurred regions leads to distorted flow fields and structural artifacts.
Aligning Through Deformable Convolutions
To solve this, researchers like Zhen Liu have moved away from rigid alignment. In ADNet: Attention-Guided Deformable Convolutional Network for High Dynamic Range Imaging, the authors use deformable convolutions to align frames in the feature space rather than the pixel space.
Instead of forcing a pixel at $(x,y)$ to move to $(x+dx, y+dy)$, deformable convolutions allow the network to adaptively sample from a neighborhood of pixels. This flexibility allows the model to "gather" relevant information from across the temporal sequence even when motion is non-rigid or complex. By combining this with spatial attention, the network can prioritize well-exposed regions while ignoring saturated or misaligned "ghost" pixels.
Hardware Assisted Temporal Anchoring
Another direction involves changing how we capture the data itself to minimize the temporal gap.
- Event-Guided Fusion: Jin Han and colleagues introduced HDRev-Net, which uses event cameras—sensors that detect microsecond-level changes in brightness—to guide the reconstruction. Because event data is nearly continuous in time, it provides a "bridge" that tells the network exactly how an object moved between two LDR frames.
- Dual-Camera Systems: Qianyu Zhang proposed Capturing Stable HDR Videos Using a Dual-Camera System, which uses one camera to maintain a consistent "anchor" stream while a second camera alternates exposures. This decoupling ensures there is always a temporally consistent reference to prevent flickering.
- Computational Filters: In a clever "optical hack," Mushfiqur Rouf and Wolfgang Heidrich demonstrated in Glare encoding of high dynamic range images that highlight information can be physically encoded into glare patterns using a star filter. This allows HDR reconstruction from a single shot, theoretically eliminating the ghosting problem entirely by removing the need for multiple exposures.
Learning Without Ground Truth
A major bottleneck in this field is the lack of "clean" ground-truth HDR data for dynamic scenes. It is nearly impossible to capture a perfectly aligned, noise-free HDR image of a moving subject.
To bypass this, Chuan Wang developed UPHDR-GAN, which uses Generative Adversarial Networks to learn from unpaired data. Instead of seeing "Before" and "After" pairs, the model learns the general distribution of what a high-quality HDR image looks like and uses a "min-patch" module to identify and suppress ghosting artifacts during training.
Similarly, Michal Nazarczuk explored Self-supervised HDR Imaging, where the model generates its own "pseudo-labels" by identifying static, well-exposed regions within a sequence. This allows the network to learn the physics of light and motion directly from raw, imperfect video footage.
Efficiency and Generative Priors
As the field matures, the focus is shifting toward real-time performance and perceptual quality. Tao Hu 0013 recently introduced GMODiff, which reformulates HDR fusion as a "Gain Map" refinement task. By using a single-step diffusion prior, the model can "hallucinate" missing details in saturated regions with the high perceptual quality of generative models, but at a fraction of the computational cost of traditional iterative diffusion.