The importance of choosing the correct scanning parameters and compression technique for initial image capture led to a series of experiments illustrating the effects of the various options, and to the creation of the decision tree (found in Appendix B) which can be used to determine the proper method for scanning a collection of known characteristics. The decision tree also assists in identifying the exceptional documents in the collection which require different choice of scanning parameters or compression, indicating how these exceptions should be treated.
For these experiments, a small number of images were used to evaluate a variety of alternative spatial and tonal resolutions. This smaller number of images allows easy comparisons when evaluating the results. Printouts of the manipulated images are included with the original copies of this report in the large binders in the Preservation Directorate Office. It is these binders which are referred to as Appendix A in this section. If you are reading a distribution copy of this report, photocopies of cropped portions of some of these printouts are found in a surrogate Appendix A. Those images will of course be degraded from the originals, but the relative quality among sets of alternative images should still be visible. Please note that since the entire images are not present in the cropped versions, some features are referred to which can only be seen in the original Appendix A (i.e. the binders of complete printouts), or in the electronic image files.
The printer used for making the printouts in the Appendix A was a 600 dpi laser printer with halftoning performed by the printer when printing gray-scale images. A printer capable of true gray-scale printing (e.g. one using a dry silver process or a dye sublimation process) would not have required the halftoning, resulting in more accurate printouts. These printers will be more common in the future, but for now, high resolution laser printers are more common, faster and less expensive.
A duplicate recordable CD-ROM disk containing all of the images is available for a pre-paid $50.00 duplication and shipping charge from:
Picture Elements Inc.
777 Panoramic Way
Berkeley, CA 94704
Spatial Resolution - Minimum Spatial
Resolution - Magnified Scans of the Engraving
Magnified Scans Type - Tonal Resolution: Grayscale and Binary
Low Contrast Bibliographic Card - Manuscript Page
Carbon Copy - Closing Note
In choosing a scanning resolution for the creation of a preservation image, the range of available alternatives is part of the problem. If there were only one alternative (and it was adequate to the task), the choice would be trivial.
There are choices to be made for both the spatial resolution (samples per inch) and the tonal resolution (bits per pixel). We assume here that the resolutions will be constant within any given scan, but they may be different from scan-to-scan based on differing characteristics of the source material. This means that if a single page has multiple types of content, the scanning resolutions chosen will have to perform well for all of the content types. When there are multiple alternatives, we will attempt to make the easy choices first, narrowing down the remaining alternatives as we go.
Monochrome document scanners commonly offer between 100 and 1200 dpi of resolution and between 1 and 4096 gray tones (1 to 12 bits per pixel).
What is surprising is that, depending on the source material, the optimum choice can be anywhere within the range (e.g. a continuous tone photograph looks much better at 100 dpi and 12 bits/pixel than at 1200 dpi with 1 bit/pixel, while 4 point type looks much better at 1200 dpi and 1 bit/pixel than at 100 dpi with 12 bits/pixel).
The other consideration is that either of these two extreme choices would produce a small compressed file size, but combining the 1200 dpi resolution with 12 bits/pixel produces a huge file size. Increasing the resolution by a factor of 12 on each axis increases the number of pixels by a factor of 144, and going to 12 bits/pixel requires a gray scale compression algorithm to be used, which is less effective at this resolution than a binary compression algorithm.
When people think of resolution, they normally think of spatial resolution first, and may not think consciously of tonal resolution at all. In accordance with this, we will start with some experiments with various spatial resolutions, but all with 256 levels of gray-scale (essentially continuous).
These experiments will show that there is a minimum spatial resolution which is necessary to eliminate visually distracting jagged edges caused by spatial quantization in the image. The experiments show that beyond this spatial resolution there is a diminishing return. Doubling the spatial resolution on each axis (e.g. from 300 dpi to 600 dpi) quadruples the number of pixels, but may not produce a visible image quality improvement when the spatial and tonal resolutions are already sufficient.
The original scanned source image shown in Figure 1 bio01l.tif in Appendix A will be used for these experiments. The printed page that was scanned to make this image was bitonal (black ink on white paper). The 8 point type is typical, but the engraved picture is challenging because it relies on subtle variations in the engraved stroke widths to achieve its visual quality.
Original source scans at 600 dpi, 256 level (8 bit) grayscale scan were used in anticipation that this would provide adequate image quality for these experiments (all other images were derived from these original scanned images).
Zooming up the images is a way to avoid the printer limitations when evaluating scanned image quality. Figure 2 bio_1.tif shows the engraving, as scanned at 600 dpi, magnified by approximately 4 times. This magnification makes it easier to see the detailed scanning quality that is achieved -- without requiring a magnifying lens. The magnified image looks better than the unmagnified image at both a 1.5 foot and a 6 foot viewing distance, indicating that it successfully avoids most of the printer-induced degradation in the unmagnified image.
In the magnified 600 dpi image (Figure 2 bio_1.tif), the strokes are all quite smooth, without any significant jags due to spatial quantization. In the various images that are scaled down to 300 dpi, namely,
there is some jaggedness visible in the engraving strokes (such as in the beard) and along the strokes of the signature at the bottom. The averaging method of scaling (Figure 7 bio1a3b.tif) most accurately represents what actual 300 dpi scanning would produce, but the other alternatives are available if the 300 dpi image is being computed. When viewed at a distance, the scaling options look nearly identical, with none of them looking quite as "crisp" as the 600 dpi image. The jaggedness is not visible when viewed at a distance.
When scaled to 200 dpi, there is significantly more jaggedness than there was at 300 dpi, which is visually quite distracting when viewed up close. In the beard, Figure 7 bio_200.tif and Figure 8 bio1a2a.tif look "pixelized", while Figure 9 bio1a2b.tif and Figure 10 bio1a2c.tif look blurry. When viewed at a distance, almost all of the jaggedness becomes imperceptible, but the impression of blurriness remains.
When scaled to 150 dpi, Figure 11 bio1_150.tif and Figure 12 bio1a15a.tif are so pixelized in the beard that the individual strokes aren't visible anymore, and Figure 13 bio1a15b.tif and Figure 14 bio1a15c.tif are blurred to the point where only some of the strokes can be distinguished. The jaggedness is now extreme enough that it can exceed a stroke width, and remains visible when viewed at a distance.
Similar results are observed with the zoomed up images of the 8-point type. The 600 dpi scanned image (Figure 15 bio2.tif) has no visible jags.
The 300 dpi scaled images have a little bit of jaggedness and blur when examined closely (see the thin stroke in the capitol "M"s), but are indistinguishable from the 600 dpi scan when viewed at a distance.
The 200 dpi scaled images have visually distracting jags when viewed up close, which are still barely visible at a distance.
The 150 dpi image has jags which can exceed a stroke width, which are clearly visible at a distance.
Microfilm Quality Index
These results can be compared with the expected Quality Index for first generation microfilm (another continuous tone technique) as presented in the AIIM TR26-1993 Technical Report entitled "Resolution as it Relates to Photographic and Electronic Imaging". Figure 4 in the report shows that 8-point type will have High Quality ("excellent, with serifs and fine detail resolved") when captured at a film resolution of 8.0 lp/mm (400 lines per inch), Medium Quality ("quite legible, but serifs and fine detail may be lost") when captured at a film resolution of 5.5 lp/mm (275 lines per inch), and only Marginal Quality when captured at a film resolution of 4.0 lp/mm (200 lines per inch).
When experimenting with tonal resolution, the two alternatives of interest are 8 bits per pixel (effectively continuous tone) and one bit per pixel ("binary"). There is no compression advantage for any intermediate tonal resolution when the compression alternatives are JPEG and Group 4.
For example, a 4 or 6 bit per pixel image, when JPEG compressed,produces essentially the same file size as an 8 bit per pixel image, so 8 bits per pixel will always be preferred when JPEG compression is being used. When the tonal resolution is reduced to one bit per pixel, Group 4 compression becomes the preferred option, since it can produce reduced file sizes.
The first experiment shows that when the tonal resolution is reduced to one bit per pixel, even the highest spatial resolution is not able to represent the fine variations in the engraving's stroke widths sufficiently well to convey the pictorial quality of the original engraving.
600 dpi Binary
In the 600 dpi binary image (Figure 16 b1_6b.tif), the strokes have a nearly uniform width, and the magnified printout looks jagged, even at 600 dpi. Scaling the image to an (artificial) 1200 dpi (Figure 17 b1_12b.tif ) removes the jags, but still doesn't provide enough variation in the stroke widths to convey brightness variations.
The 8-point type survives better when binarized at 600 dpi (Figure 18 b2_6b.tif). The contrast enhancement inherent in binarization results in solid black and bright white, making the text easy to read, but any noise in the image has also become dark black, producing distracting speckles in the image. Some jags are visible in the magnified printout.
An original 1200 dpi scan would provide some improvement, but scanners with true 1200 dpi optical resolution (such as graphic arts drum scanners) are significantly slower and more expensive than 600 dpi scanners.
300 dpi Binary
Scaling the resolution down to 300 dpi and then creating a binary image results in a very jagged magnified image, even for the signature below the engraving. When viewed at a normal magnification, the signature is clear, but the engraving is illegible.
At 300 dpi (Figure 19 b2_3b.tif) the binary image is still quite readable at normal viewing size, with some jaggedness visible. When magnified, the jags are clearly visible and distracting.
Additional experiments using a variety of documents indicate that reducing the tonal resolution runs a greater risk of dropping out important information than reducing the spatial resolution. Reducing the tonal resolution also can pick up undesired information and increase its contrast, making it more visually distracting.
Figure 20 crd3.tif shows a portion of a bibliographic card scanned at 600 dpi, with low contrast writing in the lower right corner. The 300 dpi version (Figure 21 crd3_3j.tif) looks nearly identical when viewed at normal size, even though it has gone through "lossy" JPEG compression.
Binary Used on Low Contrast
The 600 dpi binary version (Figure 22 crd3_6b.tif) has the low contrast writing picked up well due to a very sensitive threshold setting being used (we assume here that the low contrast writing is desired information and that picking it up is good).
The 300 dpi version (Figure 23 crd3_3b.tif) used a more normal threshold setting, with the low contrast writing dropping out of the image. This dropout (due to a less accurate binarizer threshold setting) is a more serious quality loss than the increased spatial quantization. Clearly, binarization places low-contrast features at serious risk, necessitating active quality control measures.
The binarization algorithm used is an extremely high quality one developed by Picture Elements. Integrated circuits implementing this algorithm are incorporated into several leading scanners. It is worth noting that more conventional binarization algorithms would lose considerably more of the low-contrast text.
Figure 24 mrt1.tif is a 600 dpi scan of a manuscript page having uneven contrast, but in which all of the words are readable.
300 dpi JPEG
The 300 dpi version (Figure 25 mrt1_3j.tif) looks nearly identical when viewed at normal size, with all of the words still readable, even though it has gone through "lossy" JPEG compression.
600 dpi Binary from 300 dpi JPEG
The 600 dpi binary version (Figure 26 mrt1_6s.tif) is readable, but noisy, due to the sensitive threshold setting.
Note: This image was scaled to 300 dpi, JPEG compressed, then scaled back to 600 dpi before thresholding, yet all of the intermediate steps have less effect on the image quality than the final thresholding operation.
300 dpi Binary
The 300 dpi binary version (Figure 27 mrt1_3b.tif) was produced by scaling down from the 600 dpi grayscale image, then thresholded using a normal threshold. The lowest contrast letters are dropped out of this image, with several of the words unreadable.
Figure 28 vrs1.tif is a 600 dpi scan of a carbon copy of a poetry fragment which is very fuzzy, but all of the words are readable.
300 dpi JPEG
The 300 dpi version (Figure 29 vrs1_3j.tif) looks nearly identical when viewed at normal size, with all of the words still readable, even though it has gone through "lossy" JPEG compression.
600 dpi Binary
The 600 dpi binary version (Figure 30 vrs1_6s.tif) is readable, but noisy, due to the sensitive threshold setting. Note that this image was scaled to 300 dpi, JPEG compressed, then scaled back to 600 dpi before thresholding.
300 dpi Binary
The 300 dpi binary version (Figure 31 vrs1_3b.tif) was produced by scaling down from the 600 dpi grayscale image, then thresholded using a normal threshold. The lowest contrast letters are dropped out of this image, with the words "creamy", "dreamy", and "Not" unreadable.
There is a lot of experience within the document imaging industry with applying digital image processing technology where access quality images are sufficient (such as the 200 dpi binary images used in fax machines), but there is little experience with the improved resolution that preservation quality requires.
These experiments have shown that there is a minimum spatial resolution which is necessary to eliminate the visually distracting jagged edges caused by spatial quantization in the image, but that beyond this spatial resolution there is a diminishing return, after which it is more beneficial to provide good tonal resolution -- even for bitonal source materials.