Daniel Reetz, the founder of the DIY Book Scanner community, has recently started making videos of prototyping and shop tips. If you are tinkering with a book scanner (or any other project) in your home shop, these tips will come in handy. https://www.youtube.com/channel/UCn0gq8 ... g_8K1nfInQ

Scan Tailor "Experimental"

Scan Tailor specific announcements, releases, workflows, tips, etc. NO FEATURE REQUESTS IN THIS FORUM, please.
xerum
Posts: 41
Joined: 12 Jul 2015, 04:23
Number of books owned: 0
Country: australia

Re: Scan Tailor "Experimental"

Post by xerum » 20 Aug 2015, 07:51

with an I7-3770 CPU and Nvidia GTX465 (picked up for $120)
16GB RAM and Win7-64bit

918 scanned pages at 300 dpi
912 of those pages were 8-bit color
6 of those pages were 24-bit color

After setting up stages 1,2, 3 and 5
note: i only used deskew (another time i will compare curved lines function)

Here is more comparison benchmarks STE vs ST9.11.1 (both 64bit)
for stages 4 (select content) and stage 6 (output = 300dpi B&W) here are the results

Stage4 / STE = 1min 50sec / ST911 = 8min 05sec / i.e. 4.4x faster
Stage6 / STE = 2min 18sec / ST911 = 19min 12 sec / i.e. 8.3x faster

When inclined i will benchmark the curved line feature.

d14b0ll0s
Posts: 31
Joined: 17 Aug 2015, 19:37
Number of books owned: 3000
Country: Poland

Re: Scan Tailor "Experimental"

Post by d14b0ll0s » 20 Aug 2015, 08:36

I think this should prove to be a nice portable machine for image processing: http://www.amazon.com/Gigabyte-i7-4710H ... op?ie=UTF8

Meanwhile, I'll see tomorrow how the quad-core Mini deals with STE if I can get it running one way or another.
Obviously, I won't be getting your speeds, xerum, since I rely on 'curves' deskewing.

xerum
Posts: 41
Joined: 12 Jul 2015, 04:23
Number of books owned: 0
Country: australia

Re: Scan Tailor "Experimental"

Post by xerum » 20 Aug 2015, 08:45

d14b0ll0s wrote:I think this should prove to be a nice portable machine for image processing: http://www.amazon.com/Gigabyte-i7-4710H ... op?ie=UTF8
.
Both CPU and Graphics on-board are faster than what i've got. I'll check some benchmarking website and let you know but for the price looks good.

Later i'll attempt benchmark comparison with curved lines and let you know.

xerum
Posts: 41
Joined: 12 Jul 2015, 04:23
Number of books owned: 0
Country: australia

Re: Scan Tailor "Experimental"

Post by xerum » 20 Aug 2015, 08:53

the link will give you an idea

my current CPU vs the machine you are looking at on Amazon

www.cpubenchmark.net/compare.php?cmp[]=2243&cmp[]=2

my cpu is just a tad faster

www.videocardbenchmark.net/compare.php? ... 1&cmp[]=82

however your amazon GTX760 craps all over my GTX465. So i would hope you would have significant gains.

d14b0ll0s
Posts: 31
Joined: 17 Aug 2015, 19:37
Number of books owned: 3000
Country: Poland

Re: Scan Tailor "Experimental"

Post by d14b0ll0s » 20 Aug 2015, 09:29

Thanks, xerum. Although the links are broken, I was able to get similar results to what you describe through gpuboss.com and cpuboss.com. At this point, I'll see if i7-3615QM brings any significant improvement first (it should).
I also tried processing my 14Mpix (4252x3264) photos with 1x, 1,5x and 2x output now, and it seems 1,5x will be the sweet spot for OCR in Adobe XI (ClearScan). I defaulted to 2x last time, so I was wrong in claiming the speed gain was only <20%, since 1,5x is closer to what I had been doing in ST911, sorry about that (I would start with 400-450 DPI input and have 600 DPI output before).
OCR is perfect in neither spot, but the final PDF quality is definitely best with 2x TIFF output (29 Mpix files), and -- somewhat surprisingly -- file sizes for OCR'ed PDFs are as follows: 1546KB, 1603KB, 1496KB for the small sample I tried, so the largest with 1,5x, the smallest with 2x. I will look at the speeds on Mini, and I might actually consider using 2x output considering the speed gain, since I work with those output PDFs a lot and rely on their quality and 'searchability'.

xerum
Posts: 41
Joined: 12 Jul 2015, 04:23
Number of books owned: 0
Country: australia

Re: Scan Tailor "Experimental"

Post by xerum » 20 Aug 2015, 09:38

benchmarking only the Curved lines function in STE.

100 samples pages from the previous 918 pages benchmarked in previous post

Open CL checked and selected GTX465 = 1min 38sec
Open CL checked and CPU i7-3770 selected = 6min 15sec
Open CL unchecked = 3min 05sec

So d14b0ll0s is correct. i.e. leaving Open CL unchecked is better performance than selecting CPU.
However a fast performing graphics card helps immensely if you have one.

Due to the curved line feature in different stages in ST911 it was difficult to gauge performance benchmark between the two. FWIW ST911 output at 300dpi with curved switched on for all 100 samples pages took 5min 08sec (if that means anything)

xerum
Posts: 41
Joined: 12 Jul 2015, 04:23
Number of books owned: 0
Country: australia

Re: Scan Tailor "Experimental"

Post by xerum » 20 Aug 2015, 09:42

d14b0ll0s wrote:-- file sizes for OCR'ed PDFs are as follows: 1546KB, 1603KB, 1496KB for the small sample I tried, so the largest with 1,5x, the smallest with 2x. I will look at the speeds on Mini, and I might actually consider using 2x output considering the speed gain, since I work with those output PDFs a lot and rely on their quality and 'searchability'.
Thanks d14b0ll0s - this is interesting. I found the larger DPI i scan and using Adobe clearscan to OCR just resulted in larger PDF searchable files with no real visual quality gain. This why i decided that 300dpi was the sweet point. It seems interesting that the higher your DPI and using 2x in STE resulted in lower PDF size.

d14b0ll0s
Posts: 31
Joined: 17 Aug 2015, 19:37
Number of books owned: 3000
Country: Poland

Re: Scan Tailor "Experimental"

Post by d14b0ll0s » 20 Aug 2015, 09:54

Thanks for these results, xerum. I think the way to compare curved lines deskewing times in ST911 and STE is to add the time needed to complete stages 3 and 6 in STE and compare it against ST911's stage 6 with deskewing on. Or actually stage 4 in both scenarios as well, since you need this to get to stage 6. This makes more sense since in 99% cases the user goes through all stages in ST (because s/he wants to see the output file).
I'm getting used to how STE handles curved lines deskewing now, and I'm finding the accuracy superior to what I got accustomed to in ST911. Well done, Tulon!

xerum
Posts: 41
Joined: 12 Jul 2015, 04:23
Number of books owned: 0
Country: australia

Re: Scan Tailor "Experimental"

Post by xerum » 20 Aug 2015, 10:07

d14b0ll0s wrote: I think the way to compare curved lines deskewing times in ST911 and STE is to add the time needed to complete stages 3 and 6 in STE and compare it against ST911's stage 6 with deskewing on.
i ok i redid the above comparison using as you suggested stage 3,4 & 6 and the result is 1 min:56 sec
this is much better than 5min+ in ST911

d14b0ll0s
Posts: 31
Joined: 17 Aug 2015, 19:37
Number of books owned: 3000
Country: Poland

Re: Scan Tailor "Experimental"

Post by d14b0ll0s » 20 Aug 2015, 10:12

xerum wrote:ok i redid the above comparison using as you suggested stage 3,4 & 6 and the result is 1 min:56 sec
this is much better than 5min+ in ST911
thanks for checking; so with CPU multi-thread support only (w/o GPU) this would be somewhere in 25-30% gain range, according to the numbers from your previous post; this is quite in line with my results (which is interesting, since you have a quad-core CPU); I'll try to install ST911 back soon and check my numbers again, but I think it should be roughly the same, since I think my earlier 15-20% number was for STE 2x output compared with ST911 ~1,5x

Post Reply