5. Preservation-quality images: types and resolution
Note: This and following sections are illustrated with selected images from the project; Appendix A includes the sample-image set. Each sample is presented in the format employed during the committee's comparative examination, including many TIFF images. In order to view the images, readers of this report must equip themselves with suitable image-viewing software. The grayscale images have a tonal resolution of 8 bits per pixel while the color images are 24 bits deep. For display of the full image quality, the viewer's display monitor adapter must accommodate these levels of tonal resolution.
Tonal versus binary images. The committee agreed with the consultants that, generally speaking, the tonal images reproduced the documents with greater fidelity than binary images. Some binary images offered equal but not superior legibility. Binary images provide significantly smaller file size and easier printability but the committee felt that these advantages were more significant in the consideration of options for access images (see Section 9 below).
Marks or strokes with different densities (thickness, darkness) could be distinguished better in the tonal images than in the binary. For example, in a typewritten script with pencil strike-overs, the underlying typed words could be seen more clearly in the grayscale images than in the binary. A high degree of verisimilitude was also noted for a handwritten document. Greater legibility was noted in the grayscale version of a printed-and-written call slip.
Note: The grayscale examples reproduced here have been contrast stretched and compressed with the JPEG algorithm; comparisons of uncompressed and compressed images are provided in Section 8 below. These are the full-resolution images. Appendix A offers more network-ready alternatives, including thumbnails.
Color images indicated both the difference in color and the actual color in the document with red-ribbon and black-ribbon typing. Grayscale images successfully indicated the differences between red and black, albeit without a specific indication of the color. In the binary images, the distinction was lost.
Based on these and other similar examples, the committee endorsed the consultant's proposal to produce tonal images when the testbed was created during the project's Phase II and to capture in color those documents having significant color content.
Levels of spatial resolution. The tonal images were presented at three levels of spatial resolution. Although the collection is too large and varied to permit an exhaustive survey, the selected documents were thought to include the smallest typical features, thus offering a reasonable test of spatial resolution. Upon examination of the images, the committee was surprised to see that the differences between the levels of resolution were not terribly significant. Although 300 dpi provided better detail than 200 dpi, few felt that increasing the spatial resolution of the tonal images to 600 dpi produced noticeable improvement for Federal Theatre Project collection documents.
Note: The 600 dpi sample was scanned at an optical resolution of 400 dpi and interpolated to yield 600 dpi. The 200 and 300 dpi samples were averaged from the 600 dpi sample.
The consultants pointed out that spatial resolution beyond the minimum necessary is costly (increasing the uncompressed image size as the square of the increase) and offers minimal gain. This is because that minimum necessary resolution preserves all the required features to assure perceived fidelity. When spatial resolution is such that one or two full pixels fall across the width of the thinnest stroke on the manuscript, that stroke will be preserved. At standard viewing distances, further increases in, for example, the accuracy of edge placement (which would come with increasing resolution) would not be seen.
The consultants also noted that an argument could be made that future processing might benefit from the increased resolution. It is worth noting, however, that in cases where a scanner delivers data at a spatial resolution above its native optical resolution (as determined by the spacing and number of the photosites in its image sensor), the future processing could accomplish this same function. Many current scanners tend to be limited to 300 or 400 dpi optical resolution; thus one could wait until a later moment to interpolate the image information to higher resolution. Successful future processing for increasing resolution (by interpolation), however, would depend upon keeping the artifacts produced by compression to a minimum; see Section 8 on compression and JPEG artifacts.
Bitonal images have a more pressing need of spatial resolution beyond that indicated by the 1 to 2 pixel per stroke dictum. When a character (whose stroke thickness varies in a smooth way, reflecting the aesthetic intent of the typeface designer) is represented by binary pixels, subtle changes in the original stroke thickness are rendered as quantum jumps in thickness: from 1 to 2 pixels (a 100 percent increase), or from 2 to 3 pixels (a 50 percent increase). Moving beyond 1 to 2 pixels per stroke lessens this effect, hence the requirement for 600 dpi in preservation bitonal images of text suggested in the widely-quoted Cornell work.
Tonal images instead lessen this quantum effect by subtly "graying out" pixels which are only partially on the stroke; they do this without an increase in spatial resolution. While this "anti-aliasing" technique has been in use in phototypesetting and vector computer graphics for 30 years, imaging designers have recently re-discovered it. When used to scale down a high spatial resolution binary image to a lower resolution tonal image for screen display, it is often called "scale to gray."