SIGGRAPH 2005 Computational photography
Removing Photography Artifacts Using Gradient Projection and Flash-Exposure Sampling
The idea
Combine two exposures.
Separate the reflections.
A flash/no-flash pair provides complementary views of the same scene. Gradient projection uses the approximate stability of gradient direction under illumination changes to reduce unwanted reflections and combine the useful information in the two images.
- Ambient image
- Captures the scene with unwanted reflections
- Flash image
- Short exposure reduces reflections but may add a hotspot
- Gradient projection
- Constrain gradient direction using the reference image
- Layer separation
- Integrate projected and residual gradient fields
Gradient direction
as a cue
Photographing through glass can capture reflections of objects in front of it. In this example, the ambient photograph reveals the photographer’s face, hands, and tripod. The short-exposure flash photograph suppresses those reflections but adds a hotspot from the glossy painting surface.
In a patch without reflections, the flash and ambient image gradients point in the same direction. Where reflections overlap the scene, the ambient gradient direction is perturbed. The method uses gradient direction as a quasi-invariant under illumination changes.
Project and
integrate
Constrain the ambient-image gradient direction by projecting each ambient gradient vector onto the corresponding flash-image gradient vector. A 2D integration of the projected field produces the result image.
The gradients removed by projection form a residual field. Integrating that residual gives the reflection layer, decomposing the original image into a reflection-reduced image and the removed reflections.
Why projection?
Knowing a gradient’s desired direction still leaves its magnitude unspecified. Projection selects a component in that direction while removing the orthogonal component.
For an integrable gradient field, rotating every gradient vector by 90° produces a divergence-free field. Reconstruction from that field has no image structure, up to the choice of boundary conditions and additive constant.
The illustrated experiment rotates gradients computed from the Lena image and reconstructs an image at each angle. At 90°, visual information disappears; at 180°, the reconstruction is the negative of the original image.
Related research
For a more general algorithm based on gradient-field transformation, see Edge Suppression Using Cross-Projection Tensors · CVPR 2006.