"Ideal" may be hard to define for image size. The wiki (I believe) says the lower case letters (for English) should be at least 20 to 30 pixels in height. By default, I scale everything by a factor of 3. If your screen is set to 96 dpi resolution, 300 dpi would be about a factor of 3. If your font size is large enough, then sometimes you can get better results without scaling, since scaling often blurs the image a little.
What I said about leptonica is for software developers building a front end to tesseract. If you are using ImageMagick, I suspect that is fine. I think 8-bit per color is standard for tesseract if you are not doing black and white. ClearType is an implementation of sub-pixel rendering, which is designed for an LCD screen with the red, green and blue sub-pixels in separate locations. Printers and scanners and OCR typically are not oriented to sub-pixels. I think OCR accuracy is better with sub-pixel rendering disabled. On Sep 8, 4:56 am, haratron <[email protected]> wrote: > I'm also interested in this topic. > > I have a couple of questions: > 1. How can I calculate the ideal image size (300dpi?) to feed to > tesseract? I mean, how do I identify how much scaling the image needs, > before the OCR procedure. > 2. I'm currently using ImageMagick's convert program for scaling and > converting to grayscale. Would it make a difference if I used > leptonica instead? > 3. Do the bits of color matter? Is there an optimal color depth? > 4. Does the OCR work best when ClearType is enabled or disabled? > > > > On Tue, Sep 7, 2010 at 11:30 PM, SteveP <[email protected]> wrote: > > Hi Quan, > > There is more than one way to scale as you may know. I have seen > > OCR fail in some cases depending on how you scale. I have a front end > > I use for my software that calls tesseract. I ended up providing > > options for scaling and options for converting from 24-bit color to > > gray or black and white. > > > Let me start with some simple answers, though. Scaling with > > interpolation seems to work best most of the time. Converting to gray- > > scale seems to work most of the time. (I read that Ray Smith did not > > design tesseract for color screen images, so I really have not > > experimented with leaving things in color.) I do not think tesseract > > pays attention to the Alpha channel since it does not pertain to when > > a single image sits by itself. (Converting to gray-scale does not > > work in general if the text is rendered with ClearType or sub-pixel > > rendering. If anybody figures out a good approach for OCR of > > ClearType, I would appreciate getting an email since I don't read a > > lot of the posts. Post your answer too.) > > > I think the scaling software at the leptonica web site is good. I > > have had some trouble with the method in Windows that uses > > createGraphics and drawImage. (Someone I worked with used the Windows > > method on a blank image and got non-blank OCR results because the > > Windows method seemed to me to introduce a row of black around a > > couple of the edges. That's how it appeared to me, but it is possible > > I did something wrong.) > > > Relative to scaling, I made a post in August about using nearest- > > neighbor scaling when the characters are close together. This is > > because scaling with interpolation without sharpening tends to blur > > the edges of text characters. Leptonica has code for sharpening, I > > believe, but I have not used it yet. Scaling by a factor of 2 without > > interpolation and then by a variable factor with interpolation to the > > needed size is a simple way to get some sharpening and some separation > > between characters. > > > On Aug 31, 7:26 pm, Quan Nguyen <[email protected]> wrote: > >> Hi Ian, > > >> I'm implementing a feature in my program to enable OCR of screenshots. > >> The results have been generally better after the captured images were > >> rescaled from 96 DPI to 300 DPI. I was wondering if other simple > >> manipulations could be done programmatically to the images to produce > >> even better results. > > >> The types of the screenshots are either 32bppArgb or 24bppRgb. Would > >> changing to grayscale or stripping the Alpha help? > > >> Quan > > >> On Aug 31, 12:17 pm, "Ian Ozsvald (A.I. Cookbook)" > > >> <[email protected]> wrote: > >> > Hi Quan. > > >> > I've used tesseract to OCR frames from 640x480 screencast videos, > >> > generally it worked > >> > fine:http://ianozsvald.com/2010/05/17/extracting-keyword-text-from-screenc... > > >> > What problems are you seeing when you try tesseract? > > >> > Ian. > > >> > On 30 August 2010 23:46, Quan Nguyen <[email protected]> wrote: > > >> > > I understand the resolutions of screenshots are typically inadequate > >> > > for OCR, but besides rescaling to a higher resolution, say, 300 DPI, > >> > > what other preprocessing operations may be needed on the images to > >> > > yield optimal OCR results? > > >> > > Thanks. > > >> > > -- > >> > > You received this message because you are subscribed to the Google > >> > > Groups "tesseract-ocr" group. > >> > > To post to this group, send email to [email protected]. > >> > > To unsubscribe from this group, send email to > >> > > [email protected]. > >> > > For more options, visit this group > >> > > athttp://groups.google.com/group/tesseract-ocr?hl=en. > > >> > -- > >> > Ian Ozsvald (A.I. researcher, screencaster) > >> > [email protected] > > >> >http://IanOzsvald.comhttp://MorConsulting.com/http://blog.AICookbook....Hide > >> > quoted text - > > >> - Show quoted text - > > > -- > > You received this message because you are subscribed to the Google Groups > > "tesseract-ocr" group. > > To post to this group, send email to [email protected]. > > To unsubscribe from this group, send email to > > [email protected]. > > For more options, visit this group > > athttp://groups.google.com/group/tesseract-ocr?hl=en.- Hide quoted text - > > - Show quoted text - -- You received this message because you are subscribed to the Google Groups "tesseract-ocr" group. To post to this group, send email to [email protected]. To unsubscribe from this group, send email to [email protected]. For more options, visit this group at http://groups.google.com/group/tesseract-ocr?hl=en.
