| bayer filtering vs foveon |
|
An image quality study based on SD9 samples by Mike Chaney many people try to get a handle on the image quality differences between Foveon's full color image sensor and traditional Bayer sensors. Since the SD9 captures all three color channels at each photo site, it is possible to "Bayerize" a few samples by discarding all but one color at each pixel to create a standard Bayer mosaic pattern. That pattern could then be "fed" to Qimage Pro in the form of a raw Bayer image, allowing Qimage Pro to recreate the image as if it were shot on a Bayer camera. In doing so, we can effectively compare a 3.4 MP image from a Foveon X3 sensor to what we would expect from a 3.4 MP Bayer sensor. Since Bayer sensors capture images very differently from full color sensors like the Foveon X3, it is important for viewers to understand the differences between the two technologies in order to understand how to evaluate images once reviews of full color sensors start to appear on digicam review sites. Since Bayer and full color sensors differ so significantly in their ability to capture different types of detail in an image, it is important to understand that comparing Bayer sensor images with full color sensor images requires examination of more than just details on a black and white resolution chart! Hopefully this page (http://www.ddisoftware.com/reviews/sd9-v-bayer/)will illustrate what to look for when comparing image resolution and quality across different sensor designs. Until now, the methods of comparing image quality from camera to camera were less important because all cameras used the same type of single color Bayer sensor. Now that full color sensors are making their way onto the market, we need to understand what we've been missing (using single color sensors) and what to expect with full color sensors. A note about validity and practical applications Please note that the process that I am calling "Bayerization" involves manipulating data from a full color device to produce a theoretical Bayer sensor. We can do this (in theory) since a Bayer sensor is analogous to a full color sensor with 2 of 3 color channels omitted at each photo site. Due to differences in the optimization of each design however, actual/practical results will vary from the examples on this page, but should generally lie somewhere between the two Bayerized versions displayed. More information regarding the difference between the Bayerized samples and Bayerized samples with AA filter is given below. In writing to Phil Askey about the information on this page, he had reservations about the actual comparisons regarding my Bayerized samples that did not include at least a simulated AA filter. AA filters (short for antialiasing filters), A.K.A. "low pass filters" are used in nearly all Bayer cameras. The AA filter is basically a "blurring" filter that is placed somewhere between the lens and the image sensor (usually right on top of the sensor). This blurring filter is used to eliminate aliasing of high frequency spacial detail due to the lens being able to resolve detail higher than the Nyquist frequency, which can result in color aliasing. This blurring filter is needed basically to reduce color distortions in the Bayer design due to its single color capture per pixel. As a result, and due to popular request, I have updated this page to include samples that simulate what a normal Bayer sensor would see if it were equipped with an antialiasing filter such as those found in most dSLR's. With this information in mind, here is a description of the samples presented on this page: In the end, actual tests of the SD9 camera will be the only true way to evaluate image quality. After comparing the images on this page, you may now be armed with more knowledge about what to look for and how to compare images. As you can see, there's a lot more to quality than just resolving black and white horizontal/vertical lines. . While this is not a complete proof of concept, my techniques, at least indicates that my Bayer-with-AA samples do a reasonable job of simulating an actual Bayer sensor with respect to detail and sharpness. Conclusion I'll leave any final conclusions regarding "resolution equivalents" or other absolute comparisons to the reader, as they really have little merit anyway until actual controlled tests can be performed. As for the samples on this page, simply download versions and zoom/examine side by side. Frankly, I was surprised at how well the image reconstructed after being Bayerized with no AA filter, however, since nearly all single color [sensor] cameras have an AA filter over the CCD, the samples that include the simulated AA filter will probably be closer to what you can expect from actual Bayer based cameras. Overall, the images held up reasonably well considering the Bayerized image started with only 1/3 the amount of captured information, however, there are some obvious areas where the Foveon X3 (full color sensor) technology excels, particularly in areas of high frequency detail. Some notable areas are shown below. The first two examples show some loss of high frequency detail in the Bayerized version while the third example shows loss of chrominance (color) detail in the parking lot gate as the Bayer interpolation algorithm "removed" color information because the sampling frequency was too low to obtain true color. Due to the increased low level detail and lack of artifacts, it is likely that the Foveon sensor design will produce images that can withstand significantly more processing, including the ability to resample images to larger sizes and produce large prints. In conclusion, some things to look for when comparing Bayer and Foveon sensor images: * Bayer sensors have a very significantly reduced resolution when resolving detail comprised of mainly red/blue primaries, such as a red sports car with black pin stripes, a blue sweater with red lettering, red soda can with black lettering, etc. In these cases, resolution of the Bayer sensor is reduced to less than 1/4 of its "image" resolution! Black and white details will show the highest resolving power on a Bayer sensor, while saturated color detail will vary greatly. A Foveon sensor is much more consistent, resolving near the full resolution of the images for every color combination. * Bayer sensors will produce images that are softer and less detailed due to the "smoothing" needed to eliminate artifacts and color distortions. * Bayer sensors tend to omit chrominance (color) information when sampling high frequency detail. If you look at a picture of a tree that has many small branches with a brick wall behind it for example, you will see that many of the smaller branches "morph" into the color of the bricks in the background. This is because the branches are not wide enough to cover the multiple pixels needed to derive accurate color information on a Bayer sensor. Full color sensors completely eliminate this problem. * Bayer sensors tend to produce color moire on high frequency detail like the cuff of the jeans below (more visible in center image). |