I just have a fundamental question while studying image deblurring.
Many researchers have studied non-blind deblurring under the assumption that the blur kernel is known, but I am fundamentally curious about how they know this blur kernel.
While looking at other references, I found that if we know the temporal information of the camera, we can know the blur kernel, but they treat it as self-evident and do not provide specific methods.
For example, if the camera moves approximately 30 pixels to the right during the exposure time, I'm curious that how we can know the blur kernel based on this.
Probably, moving camera with constant speed is equivalent to convolution of original signal with rectangular function of 30 pixels size. And in frequency domain it will be multiplication of signal spectra and spectra of convolution function.
The blur, be it defocus or motion, can be modeled as a convolution. You can in theory obtain the blur kernel by deconvolution, i.e. by taking the ratio of the blurred image and the sharp image of the same scene, in the Fourier domain.
If you look at a single black point, you directly obtain that kernel (also called the point-spread function).
In the case of motion blur, you might have the possibility to take a picture of the static scene. In the case of defocus blur, either you can synthetize the sharp scene, or you can use a better lens or refocus the lens.
- Badges
- Science topic
- Similar topics
- Geoscience
- Asia
- India