The Single Most Important Factor: Resolution is the single most important factor to consider when designing the digital image preservation system. It is critical to have an in-depth understanding of both film and digital image resolution to make informed design tradeoffs involving quality versus cost.
Designing around the amount of data: A 600-dpi image is composed of 360,000 pixels per square inch, which amounts to four times the 90,000 pixels per square inch areal density generated at 300 dpi. An 8.5 x 11 inch page in the binary format at 600 dpi resolution requires 4.207 MB of storage before compression. To get an idea of just how much data this is, it would take almost four hours to transmit one 600-dpi binary image at 2400 baud, or, it would take three double-density floppy disks to store the image. If we want a truly archival image with greyscale, multiply this figure by eight. It’s the amount of data being captured, moved, and stored that most affects the design of the preservation system.
How high resolution affects cost: A high-resolution archival system must be designed with more powerful processors, higher capacity communication channels, more random access memory (RAM), more magnetic disk storage capacity, more scanners, and possibly custom hardware to handle the compression, decompression, and processing of high-resolution images. These components increase the cost of the system. In fact, the capital costs and operating costs of a digital image preservation system are directly proportional to the resolution it is designed to handle. The resolution decision is absolutely crucial. Not only does it affect the design of the system, but it also determines the maximum possible quality for each image captured.
How Much Image Resolution Is Required? If the system’s objective is to preserve the information content of the page, it can be accomplished at a much lower scanning resolution than would be required to preserve a high-fidelity copy of the original. A resolution of 300 dpi (adequate access resolution) will preserve about 99.9 percent of the page’s information content. One need only examine a page transmitted by a typical fax machine to see that even at a resolution of 200 dpi (which is the resolution of today’s fax–see Figure 10), all but the smallest type fonts and finest lines are faithfully, albeit crudely, rendered. Most loss occurs in the area of the halftones, yet most of the intelligence in the page is preserved.
Capturing small type sizes: Let’s assume one design requirement is to be able to read footnotes from the captured page that are in four-point type. A four-point character has a height of 4/72 of an inch (each point is 1/72 of an inch). Assuming that the theoretical character is formed within a cell that is five lines high by five lines wide, each line that forms the character would be approximately 1/5 of the character’s total height, or in the case of a four-point character, the line width is 4/72 * 1/5 = 1/90 or .011 inch wide. To capture this character legibly, the scan resolution must be at least fine enough to have two or more scan lines (assume three) overlay each line that composes the character. This means that each scan line must be no greater than .0037 (.011/3) inch, which translates into a scanning resolution of about 300 dpi (1/.0037 = 270).
Figure 11a shows an enlarged portion of the IEEE test chart scanned at 300 dpi. (The IEEE chart is shown in its entirety in Figure 4.) Looking closely at the four-point type roughly in the center of the page, we can see that there are approximately 12 scan lines from the top to the bottom of the character. (The black lines are electronically generated every eighth line.) This measurement is designated the “X-height” of the character (see Figure 11b). The actual “body height” of the character is approximately 30% larger. We can calculate using the formula 4/72 * 300 dpi = 16.6 that approximately 16 scan lines cross the body height, which is also known as the point size. (Type size expressed in points does not refer to the actual dimensions of the character but to the height of the metal surface on which the raised design is produced for typesetting.) From the close-up in
Figure 11a, we can clearly see that four-point type is at the limits of the resolving capability of 300 dpi scanning. The two-point type below and to the right in Figure 11a has been completely lost at 300 dpi.
At 400 dpi (see Figure 12) the four-point type, located at the bottom of the page and slightly to the right, is much more legible. Figure 11a and Figure 12, graphically illustrate the role resolution plays in capturing a high fidelity replica of the original page.
The typeset parameters for differing groups of documents vary according to font type, document age, background coloration, and other factors. To verify the above calculations for a specific set of documents, the designer should test a random sample of the material to be scanned, along with some resolution test targets, to determine the smallest fonts that must be captured. The scanning system should provide the capability to zoom in on (magnify) a small portion of the scanned page (as we have done in Figures 11a and 12), even down to the individual character level, to determine if the characters are being properly formed. The pages should also be viewed on a full-page display (100 dpi) to verity that the reduced resolution display will be adequate for operator processing and quality control.
In addition, a greyscale test chart should be used to measure the number of levels of grey being reproduced.
Finally, the test pages should he printed to verify that the print resolution is sufficient. If halftone fidelity is important, the appropriate image enhancement processes might be required as part of the system.
Halftone resolution: Surprisingly, it takes less input resolution to produce good digital halftones than to render good quality text. A simple formula for input is: a scanning resolution of about 1.5 times the desired output screen ruling* is sufficient for scanning. To get a good 133-1ine screen (equivalent to the resolution in a typical magazine), images need only be scanned at a minimum resolution of 200 dpi with greyscale. This rule works as long as the image is printed at the same size as it was originally scanned.
On the output side: the relationship between printing resolution (line screen) and number of greyscales can be determined by the following equation: number of greylevels = (printer output resolution / line screen) squared + 1. If you try to print the same 133-1ine screen on a 300-dpi printer the result is 6 levels of gray. However, if you drop the line screen down to 50 using the same printer you get 37 greylevels [(300/50)squared + I = 371, which is about optimal for a 300 dpi laser printer. Of course, a 50-1pi diagonal screen is coarse, but as you can see from Figure 13, it renders the halftone with some degree of fidelity. If the resolution of the output device is held constant, then as screen resolution increases, the number of greyscales decreases. This is why a high resolution output device is necessary to render high screen resolution while at the same time reproducing a high number of greylevels.
Archival resolution: As stated above, binary resolution on the order of 1,000 dpi is required to create an image comparable in quality to an image stored on film. Theoretically, one should operate at resolutions as close to this as possible. But due to the high cost of doing so, it is just not practical with today’s technology. If the objective is to capture every detail of the smallest type fonts, the finest of the graphic art lines, and produce an accurate rendition of any halftones on the page, the required resolution would have to be about 600 dpi or more. Archival resolution is designed to preserve a faithful replica of each document.
Greyscale scanning can improve page quality: Regardless of the resolution, digital page image quality can be improved by using scanners that capture the image in shades of gray. The additional greyscale data can be processed electronically to sharpen edges, fill in characters, remove extraneous dirt, remove unwanted page stains or discoloration, and, in effect, create a much higher quality image than possible with binary scanning alone. A major drawback to scanning in greyscale is the large amount of data captured. Since this is the case, methods must be found that will take advantage of the additional greyscale input data to produce a high-quality image while minimizing the amount of data needed to be stored. Image enhancement is just such a method.
Image enhancement: After scanning the page with a greyscale scanner, the digital greyscale data can be mathematically manipulated to automatically decompose the page into text/ line-art areas and halftone areas and to process each area of the page with mathematical formulas (filters) that maximize content quality on each area of the page separately. For example, on the text area of the page, an edge detection filter could more clearly define the character edges, a second filter could remove the high-frequency noise (stray ink spots or dirt), and, finally, another filter could fill in characters. Greyscale areas of the page could be processed with different filters to maximize the quality of the halftone. See Figure 14.
In addition, the contrast range of digital image greyscale data can be increased–that is, the various greylevels captured during the scan can be recorded into a histogram with values from zero to 256 (if scanning at eight bits per pixel). As an example, let’s look at an original photograph, a reproduction made on a typical photocopier, an image scanned in binary mode, and a mathematically enhanced image. (See Figure 15a.) It is clear that the enhanced image comes much closer to the quality of the original than any of the other reproductions. Next we’ll view the same page after greyscale scanning, mathematical enhancement, and rehistogramming. The image at the top of Figure 15b and the graph show that most of the values captured in the original histogram are spread over a fairly narrow range of the greyscale, from zero to 100. It is easy to spread those sample points over the entire greyscale range to improve contrast and make unreadable areas of the page readable. This is done by remapping a narrow range of greyscale onto a wider range and separating the levels so there are about 30 gradations in the image area. (See the two images on the lower half of Figure 15b.) This rehistogramming technique, which is also called “stretching the gamma,” very effectively increases the quality of an image.
Finally, stains and discoloration can be removed using background filters, and the page can be restored to look much as it did when originally published. After some processing, the enhanced page image can be written to film as a greyscale image or it can be “thresholded” to remove greyscale and reinterpret the data on optical disc as an enhanced binary image.
Standard compression algoritbms: One way to reduce digital image page storage requirements is to compress data. Binary data compression is accomplished by algorithms known as the CCITT Group 111 and IV Facsimile Compression Algorithms*. They work by removing redundancy. The algorithms represent strings of either black or white pixels (run lengths) by a code. These codes are a shorthand way of representing the black ink and white space on the page. Facsimile algorithms are lossless and completely reversible–that is, the original scanned image can be re-created exactly from the compressed data. Average compression ratios of ten or 20 to one are possible, which means that an exact replica of the page can be recreated from as little as five percent of the scanned data.
The greyscale compression algorithms developed by the Joint Photographic Expert Group (JPEG) promise high-compression ratios for greyscale data. This compression works by finding areas of the page that have some common tone, shade, color, or other characteristics and representing this area by a code. But this compression is achieved at the cost of some loss of data. Preliminary testing indicates that a compression of about ten or 15 to one can be achieved without visible degradation in image quality. Since this algorithm is not completely reversible, more testing must be done before it can be used with complete confidence.
Equating scanner resolution and film resolution: Film resolution is measured in line-pairs per millimeter. By definition, one line-pair is equivalent to two digital image scan lines. To scan an original page at 600 dpi with the objective of storing that page on 16mm film at a reduction ratio of 24X, resolution could be compared as follows:
Since one inch equals about 25 millimeters and since the reduction ratio being used is 24X, then one square inch on the original will be recorded on approximately one square millimeter of the film. Given the above, a scanning resolution of 600 dpi (300 line-pairs/inch) translates into a film resolution of 300 line-pairs per millimeter. However, because of the Nyquist sampling error (see below), one third of this resolution could be lost. Therefore, for this example, the effective resolution on the film is about 200 line-pairs per millimeter. When working with a reduction ratio of 24X, the following simple rule can be used (see Figure 16):
film resolution (line-pairs per mm) = binary scanning resolution (dots per inch) /3
For 35mm film with a reduction ratio of 12X, where approximately one inch of the original page maps onto about two millimeters of the film, the resulting film image is about twice as large as the image generated in the example above. The simplified rule for 12X reduction could be stated as:
film resolution (line-pairs per mm) = binary scanning resolution (dots per inch) /6
It should be noted that this formula includes sufficient resolution to overcome the sampling error. The general formula is:
film resolution (line-pairs per mm) = (binary input scanning resolution (dpi) /2 scan lines per line-pair) * (reduction ratio/ 25.7 mm per inch) * .66 Nyquist error
The Nyquist sampling theorem: line-pairs scanned with equivalent-sized pixels have an equal probability of coming out black or white since the scan lines do not line up precisely with the black lines in the image. Therefore, a reduction in the pixel size, which is the same as doubling resolution, is needed to ensure accurate capture of image detail. This sampling error phenomenon is known as the Nyquist sampling error.
Digital Image Scanners: How They Work
Binary scanners: Most scanners available today operate by moving a light-sensitive CCD array down the page at a fixed rate. The CCD has a sufficient number of discrete sensors (CCD elements) to generate the specified number of samples per inch (resolution) in the horizontal direction, multiplied by the width of the page. For example, to sample an 8.5 inch wide page at 300 dpi, the CCD array would require a minimum of 2,550 elements. See Figure 17.
The speed of the electronics combined with the rate at which the array moves down the page governs the vertical resolution. Each CCD element records the amount of light reflected off the page as measured by the changes in their electrical charges, like a sort of thermometer. This CCD thermometer records a value, between zero and some upper limit for each dot (pixel). In binary scanners a threshold value is selected to convert this analog representation of light into a binary value of either black (0) or white (1). We can compare the threshold value to the freezing point (0¡ Centigrade) of water. On a typical Centigrade thermometer, if the temperature is above 0¡ C, water will not freeze. If the temperature is below this level, water will freeze. In the scanner, everything below the selected threshold point is defined as black (0) and everything above that point is white (1). In binary scanning, no greylevels are preserved.
Binary page storage requirements: To determine the uncompressed size of a journal page stored at Adequate Access Resolution, the formula is:
BS = L * R * W * R; where BS is binary page storage requirement (bits) L represents page length (in.) W represents page width (in.) R represents the scanning resolution (dpi)
Given an 8.5 x 11 inch page and a scanning resolution of 300 dpi, the above formula gives an uncompressed storage space of 8,415,000 bits or, dividing by eight, 1,051,875 bytes. If we assume a compression ratio of 12:1, which is typical for CCITT Group IV compression of an average journal page, the per-page storage requirements can be reduced to an average of about 90 kilobytes (KB). Book pages can be stored in about 45 KB due to their smaller size and general lack of greyscale.
Greyscale scanners: Higher quality scanners can scan greyscale–that is, they have the capability to represent the amount of light being reflected off the page at each pixel by a value recorded by the CCD element. To return to the thermometer analogy, we are now interested in storing the exact temperature represented by the reading on the CCD thermometer, not just whether the temperature is above or below freezing. The number of greylevels recorded determines the number of bits required to store each pixel. Sixteen greylevels requires four bits (two to the fourth power) to represent it. At eight bits per pixel, the scanner can represent up to 256 levels of gray. The eight bits per pixel metric is the level usually referred to when discussing high quality monochrome scanning requirements because the eight bits will allow 256 greylevels to be stored. Although studies indicate that the average person can only perceive about 30 levels of gray, capturing 256 levels provides sufficient over- sampling of the data to reconstruct at least 32 discrete greylevels.
Greyscale page storage requirement: Since in the greyscale image we are storing eight bits of data for each pixel sampled, the formulas given above for binary storage space must be multiplied by eight to give the formula for per-page greyscale storage space:
GSS = 8(L * R * W * R); where GSS is greyscale storage space requirement (bits) L represents page length (in.) W represents page width (in.) R represents the scanning resolution (dpi) 8 is the number of bits per pixel or depth of greyscale
Capturing the same 8.5 x 11 inch page in greyscale at a scanning resolution of 600 dpi and a depth of eight bits per pixel, the above formula gives an uncompressed storage space of 269,280,000 bits, or dividing by 8 = 33,660,000 bytes. Assuming a JPEG compression ratio of 15:1, which is the maximum attainable without perceptive loss of data, the average journal page captured at Archival Resolution requires compressed greyscale storage space of 2.244 Megabytes, 1.13 Mbytes for a book size page.
The laser printer: The laser printer is currently the primary output engine for digital image printing. The laser printer is used almost exclusively because interface boards are available that connect to the laser engine directly and can drive it at video rates (about one megabit per sec.), dramatically increasing print speeds. Printing a digital image through a serial or parallel printer port would take about a minute per page. Printing through the video interface takes about eight seconds a page. Other arguments for the laser printer include their high-resolution printing capability, size, convenience, and quiet operation.
Creating a halftone: In standard printing processes, pictures are formed by placing ink dots in a pattern that creates the illusion of a photograph. The resulting image is called a halftone. The thickness and spacing between the dots are constant, but the dot size varies. The screen ruling measures the frequency of halftone dots at an angle. A newspaper has a screen ruling of about 80 dpi; a medium-quality magazine about 133 dpi; and a high-quality art book might have a screen ruling of 150 to 160 dpi. Halftone dots that are closer together tend to look more like original photographs.
Halftone printing with a laser printer: The halftone printing dot is different from the greyscale scanning dot. The dot created by a greyscale scanner contains greyscale information (depth) that represents the degree of light (shade of gray) reflected from the page at that particular point on the page. However, since printers can only print black dots, a halftone printer dot is actually a group of black dots arranged in a cell that gives the illusion of a halftone.
A key objective of any imaging system is to reproduce a high- quality, high-fidelity rendition of each page image. A laser printer has a difficult time representing halftones because it synthesizes greyshades by grouping black dots together into grids or cells (sort of a super pixel) that represent the halftone dot. These cells containing certain patterns of black dots are interpreted by the eye as a halftone. (See Figures 18 and 19.) For a 300-dpi laser printer, the optimal screen ruling has been determined by testing to be about 50 of these cells (halftone dots) per inch. This gives the right balance between the coarseness of the screen pattern and the amount of greylevels this group of patterns can represent.
Finally, PC-boards are available that use techniques for modulating the printer’s laser beam to create smaller dots at more frequent intervals, thus increasing the horizontal resolution and consequently increasing the number of levels of greyscale that can be produced on the standard 300-dpi laser printer.