Slashdot Mirror
New Camera Sensor Filter Allows Twice As Much Light
bugnuts writes "Nearly all modern DSLRs use a Bayer filter to determine colors, which filters red, two greens, and a blue for each block of 4 pixels. As a result of the filtering, the pixels don't receive all the light and the pixel values must be multiplied by predetermined values (which also multiplies the noise) to normalize the differences. Panasonic developed a novel method of 'filtering' which splits the light so the photons are not absorbed, but redirected to the appropriate pixel. As a result, about twice the light reaches the sensor and almost no light is lost. Instead of RGGB, each block of 4 pixels receives Cyan, White + Red, White + Blue, and Yellow, and the RGB values can be interpolated."
"We've developed a completely new analysis method, called Babinet-BPM. Compared with the usual FDTD method, the computation speed is 325 times higher, but it only consumes 1/16 of the memory. This is the result of a three-hour calculation by the FDTD method. We achieved the same result in just 36.9 seconds."
What I don't get is calling the FDTD (finite difference time domain) analysis as the "usual" method. It is the usual method in fluid mechanics. But in computational electromagnetics finite element methods have been in use for a long time, and they beat FDTD methods hollow. The basic problem in FDTD method is that, to get more accurate results you need a finer grids. But finer grids also force you to use finer time steps. Thus if you halve the grid spacing, the computational load goes up by a factor of 16. It is known as the tyranny of the CFL condition. The finite element method in frequency domain does not have this limitation and it scales as O(N^1.5) or so. (FDTD scales by O(N^4)). It is still a beast to solve, rank deficient matrix, low condition numbers, needs a full L-U decomposition, but still, FEM wins over FDTD because of the better scaling.
The technique mentioned here seems to be a variant of boundary integral method, usually used in open domains, and multiwavelength long solution domains. I wonder if FEM can crack this problem.
sed -e 's/Chuck Norris/Rajnikant/g' joke > fact
"You ALWAYS need antialiasing when you discretize."
That's my motto!
This space available.
Interesting comments from both, but I believe you both missed the point. The real question is, which one of these methods, FDTD or FEM-FD, will allow optimal reprocessing in the frequency domain that makes my dinner look prettier with an Instagram vintage filter?
Your eyes actually aren't sensitive to red, green, and blue. Here are the spectral sensitivities of the red, green, and blue cones in your eye. The red cones are actually most sensitive to orange, green most sensitive to yellow-green, and blue most sensitive to green-blue. There's also a wide range of colors that each type of cone is sensitive to, not a single frequency. When your brain decodes this into color, it uses the combined signal it's getting from all three types of cones to figure out which color you're seeing. e.g. Green isn't just the stimulation of your green-yellow cones. It's that plus the low stimulation of your orange cones and blue-green cones in the correct ratio.
RGB being the holy trinity of color is a display phenomenon, not a sensing one. In order to be able to stimulate the entire range of colors you can perceive, it's easiest if you pick three colors which stimulate the orange cones most and the other two least (red), the green-blue cones most and the others least (blue), and the green-yellow cones most but the other two least (green). (I won't get into purple/violet - that's a long story which you can probably guess if you look at the left end of the orange cones' response curve.) You could actually pick 3 different colors as your primaries, e.g. orange, yellow, and blue. They'd just be more limited in the range of colors you can reproduce because their inability to stimulate the three types of comes semi-independently. Even if you pick non-optimal colors, it's possible to replicate the full range if you add a 4th or 5th display primary. It's just more complex and usually not economical (Panasonic I think made a TV with extra yellow primary to help bolster that portion of the spectrum).
But like your eyes, for the purposes of recording colors, you don't have to actually record red, green, and blue. You can replicate the same frequency response spectrum using photoreceptors sensitive to any 3 different colors. All that matters is that their range of sensitivity covers the full visible spectrum, and their combined response curves allow you to uniquely distinguish any single frequency of light within that range. It may involve a lot of math, but hey computational power is cheap nowadays.
It's also worth noting that real-world objects don't give off a single frequency of light. They give off a wide spectrum, which your eyes combine into the 3 signal strengths from the 3 types of cones. This is part of the reason why some objects can appear to shift relative colors as you put them under different lighting. A blue quilt with orange patches can appear to be a blue quilt with red patches under lighting with a stronger red component. The "orange" patches are actually reflecting both orange and red light. So the actual color you see is the frequency spectrum of the light source, times the frequency emission response (color reflection spectrum) of the object, convolved with the frequency response of the cones in your eyes. And when you display a picture of that object, your monitor is simply doing its best using three narrow-band frequencies to stimulate your cones in the same ratio as they were with the wide-band color of the object. So a photo can never truly replicate the appearance of an object; it can only replicate its appearance under a specific lighting condition.