
Everyone, I've placed online several bitonal test images (lossless PNG format) derived from scans of 5 point, 6 point and 11 point text. There are both 600 dpi and 300 dpi versions for each point size. http://www.openreader.org/600vs300/ (I won't go into detail of how I generated these, except to say the original scans were done at 600 dpi optical full color, then a 300 dpi full color version was generated by high-quality resampling. Then both the 600 dpi and derived 300 dpi full color were converted to bitonal by thresholding, where the threshold values were adjusted by eye to give consistent results (my eyes, for better and for worse.) Nowhere in the process was a lossy format like JPEG used. For each point size, the best way to visually compare the 600 dpi with the 300 dpi is to print them out using Photoshop or Paint Shop Pro (or similar higher-end graphics program), adjusting the scaling so the same block of text appears identical in size on paper. From this quickie experiment, which certainly could be improved, I make the following preliminary conclusions: 1) If the scans are to be used for OCR only (which is DP's focus at present), then from the visual test alone 300 dpi bitonal appears sufficient for 5 point and larger Latin character set text -- this applies to just about all text documents DP will ever encounter. Greyscale certainly improve things, but bitonal appears to be minimally sufficient for OCR. Of course, this observation, based solely on eye, jives with DP's OCR experience. So I'm not concluding anything revolutionary here. 2) Likewise, 300 dpi bitonal is *readable* by human beings for 5 points and larger Latin character set text. 3) However, for smooth, comfortable readability, 600 dpi is definitely better, even for the 11 point text. 300 dpi clearly looks ragged (especially the 5 point text). Of course, anti-aliasing during presentation will overcome some of this raggedness, but such anti-aliasing is strictly artificial and won't fix letters which are mangled in some manner due to the reduced resolution. If the purpose of scans is for multiple use cases (and not only OCR), then it appears wise to scan text at 600 dpi, preferably 24-bit full color (which aids with image cleanup, and of course necessary when we are dealing with colored text and color illustrations.) These master scans can be resized and/or reduced in color depth using batch image processing for whatever purpose is required (e.g., direct online reading, OCR, etc.) Of course, the huge downside to higher-rez, higher-color-depth images are much greater file sizes. This causes difficulty with online archival storage and transport. For the short term, these master images probably need to be stored offline on some sort of storage media (such as DVD-ROM, tape or removable high-capacity harddrives.) [see note below] At this point, since DP is not concerned with archivability and multiple use cases, and has limited bandwidth and disk storage, then there is no reason for them to require 600 dpi for scanning of text. (Illustrations are another matter, as Juliet has noted.) But I do believe that those who are submitting scans to DP should seriously consider doing all scans at 600 dpi full color (especially if the scans can be done without page distortion such as if the book is chopped and run through a sheet-feed scanner), and then resample them to 300 dpi (bitonal or greyscale) for submission to DP. Backup the original master scans in lossless form until they can be donated to a future page scan archive. Just some thoughts.. Comments? Jon Noring [Note: Obviously, high-resolution, high-color depth scans can be subtantially compressed using a lossy algorithm, such as JPEG and a few others. However, lossy compression adds artifacts to the images, so I believe lossy algorithms should be avoided for all steps in the process of producing the master scan images. Rather, use PNG or other high-quality lossless compression algorithm for all steps in producing the master scan images.]