From the entire image (left), two 300 x 400 pixel sections were copied and saved as examples of the 144 megapixel scanning back image.  These example sections were selected to contain both high and low contrast highlight and shadow detail, along with a selection of natural colors.  Copies of these two sections were then processed as described below to yield images equivalent to a 39 megapixel DSLR image, and a 16 megapixel DSLR image.  Finally, the simulated DSLR images were interpolated up to match the size of the original scanning back sections to more clearly demonstrate the differences, and moderate sharpening was applied to all of the comparison images. before saving them as high-quality JPEGs for this article.

Keep in mind that these are TINY sections of the original scanning back photo -- each section is only one-quarter of one percent of the entire image area.  If you are viewing this article on a typical computer display with around 100 pixels per inch, you are looking at the equivalent of 3 by 4 inch sections from a 60 by 80 inch print!  However, a “photographic-quality” print requires more pixels per inch than a computer display, with 300 pixels per inch being the generally accepted limit of human vision (when you’re young, anyway…).  If a photographer intends to make a truly photographic-quality print, at least 200 pixels per inch should be utilized, in which case an uninterpolated 6,000 by 8,000 pixel scanning back image would make a 30 by 40 inch print, and these on-screen examples could represent 2X magnifications of sections from a 30 by 40 inch print. 

You can download one megabyte layered TIFF files containing all three images of each section.  By turning on/off the visibility of each layer of the TIFF files, one can quickly switch among the 144, 39, and 16 megapixel images for a much clearer view of the differences.

These results might look a little different than comparisons made using a still-life subject, in part because this large format landscape photo contains considerably more fine detail than a close-up photo of a camera or doll.  Furthermore, there isn’t any difference in size, brightness, contrast, or color among these images, other than those changes introduced by the pixel processing, so it’s easier to see more subtle differences.  Finally, these images are shown at their pixel sizes described in this article, without any subsequent “resizing for publication”, as has sometimes occurred in other comparisons. 

Since the beginning of silver-based photography, using a larger camera has generally allowed the photographer to make a larger, higher-quality printed reproduction.  Digital cameras have complicated this issue somewhat by introducing a wider variety of pixel sizes and capture technologies than was previously available, often providing better performance than an equivalent film format.  With digital imaging, the amount of information being captured is not always just a function of the size of the capture device.  However, it is still safe to say that recording a greater amount of original information for each image will allow the photographer to make a larger, higher-quality printed reproduction. 

Because of the almost universal use of Bayer pattern image sensors in instant-capture digital cameras, many readers might think that the megapixel rating of a digital camera is only an expression of its pixel dimensions, since this happens to be true for the vast majority of digital cameras on the market.  However, the (former) JCIA megapixel rating specification is really intended to express the information-capturing ability of a digital camera, so three-CCD and scanning cameras are allowed to count all three colors of their pixels, even though the different-colored pixels are effectively stacked on top of each other.  The additional full-color pixel data from the scanning back represents considerably more image information than is captured by a Bayer pattern sensor – hence, the considerably higher megapixel rating for the scanning back.  

Details of the simulation process:

Simulating an instant-capture digital image from the commonly-used Bayer pattern sensor involves more than simply reducing the pixel dimensions of the larger scanning back image to those of the DSLR image.  This is because the scanning back image contains actual red, green, and blue pixel data at every spatial location in the image, whereas two-thirds of the DSLR image has been interpolated from a single matrix of red, green, and blue pixels, typically arranged as shown in the illustration below.

After image capture, each color channel of image data from a Bayer pattern sensor is usually enlarged to the full size of the array, with the missing pixels in each color channel interpolated from nearby pixels of the same color.  Aggravated by the offsets among the color channels, this interpolation often causes colored artifacts to appear in areas of high-contrast fine detail, commonly known as “color aliasing” artifacts.  Image processing is used to avoid or remove these artifacts, sometimes by putting a slight blurring filter in front of the sensor (an “anti-aliasing filter”), and often by post-processing the image data using sophisticated color noise removal algorithms. 

This demonstration uses several different image resizing methods available in Photoshop to derive the DSLR images, each where it is most appropriate:

1.  “Nearest Neighbor” interpolation – a simple method of resizing an image that first determines from where in the existing pixel array each new pixel must be derived, and then picks the nearest single pixel from the existing array to represent the value of the new pixel.  Really not “interpolation” because no new or intermediate values are generated -- just copies of the existing values. 

2.  “Bilinear” interpolation – a somewhat more complicated method of resizing an image that first determines from where in the existing pixel array each new pixel must be derived, and then calculates a new value from the values of the existing pixels around each new pixel, using linear proportioning.  Generates a smooth result using all of the original image information. 

3.  “Bicubic” interpolation – an even more complicated method of resizing an image that first determines from where in the existing pixel array each new pixel must be derived, and then calculates a new value from the values of the existing pixels around each new pixel, using nonlinear polynomial proportioning.  Can generate a sharper-looking result than Bilinear interpolation, but may not be as faithful to the original image information. 

The process for simulating lower-resolution Bayer pattern image data involves the following steps, shown here with image examples for the 16 megapixel DSLR simulation. 

1.  Make a copy of the 144 megapixel example section, and reduce the pixel dimensions of this section using Bilinear interpolation to be equal to the pixel dimensions of an equivalent section from a 16 megapixel DSLR.  (See below for the pixel dimension calculations.)  This reduced-resolution section from the scanning back still has three times as much information as a Bayer pattern DSLR image.

2.  Make a copy of each individual color channel as a new (grayscale) image, and reduce the pixel dimensions of each color channel independently to be equal to the number of pixels in each color channel of an equivalent section from a 16 megapixel DSLR.  (See below for the pixel dimension calculations.)  Use Nearest Neighbor interpolation for the red and blue channels, and Bilinear interpolation for the green channel.  To simulate the offset among color channels, rotate the blue channel 180 degrees before reducing its pixel dimensions -- because each image section contains an even number of pixels in each dimension, and each section is being reduced by exactly one-half in each dimension, Nearest Neighbor interpolation will choose different pixels for the rotated blue channel than for the un-rotated red channel.  Rotate the blue channel another 180 degrees after this reduction to restore its proper orientation.  Bilinear interpolation is used for the green channel to provide more uniformly distributed results than Nearest Neighbor interpolation would yield. 

4.  The interpolated grayscale images are recombined into a color image by copying each grayscale image and pasting it into the appropriate channel of a new RGB image.  Because of the one-pixel offset between the red and blue channels, some color aliasing artifacts will be visible along high-contrast edges, similar to those from a Bayer pattern sensor prior to post-processing.  Viewing the individual color channels at high magnification will also reveal this one-pixel offset.  Most digital camera users never see this (simulated) native, unfiltered image data. 

5.  The recombined color image is processed to remove the color artifacts by using Adobe Photoshop CS2’s new “Reduce Noise” filter, which appears to use techniques similar to those that “raw” data pre-processors use to remove color noise artifacts from instant-capture image data.  The “Reduce Noise” filter is used at its default settings, with the “Reduce Color Noise” setting having the greatest effect on removing the color aliasing artifacts.  Applying this filter magically removes nearly every trace of color artifacts, but also has a dulling effect on other small colored areas, similar to what is observed in a processed DSLR image.  Interestingly, viewing the individual color channels at high magnification after applying this filter appears to have eliminated the offset among color channels.

6.  Finally, the interpolated and filtered image data would typically be sharpened by one or more methods to provide improved apparent sharpness in the final image.  In this case, we are ultimately going to interpolate the image data back up to the size of the 144 megapixel scanning back image for comparison, and there is no shortage of debate about the best methods for enlarging and re-sharpening a digital image file.  For this comparison, a moderate amount of Unsharp Masking is applied to the filtered color image data, the sharpened image is then resized using Bicubic interpolation, and the interpolated image has another moderate amount of Unsharp Masking applied.  Our objective was to tidy up the image as much as possible without producing obvious sharpening artifacts.  Just the same, it’s hard to avoid some signs of distress in the final interpolated 16 megapixel images -- the original image sensor data has been magnified up to 320% by the time we are finished.