[X] My Assistant

We have an "ask a question and seek help" threads for translators and scanners, but nothing for editors. This thread is aiming to correct this oversight. Have a question about editing that is not answered in our FAQs? Or perhaps you just want to show a page or two and ask for comments? Do it here. Enjoy,

This post has been edited by Red_Piotrus: Feb 10 2014, 01:52

@Izur
I thought of that but I wanted to retain the feel from the raw (IMG:[invalid] style_emoticons/default/smile.gif)

@Dammon
Yeah I might do that although it's not the cleanest way, thanks (IMG:[invalid] style_emoticons/default/smile.gif)

@Dammon
Yeah I might do that although it's not the cleanest way, thanks (IMG:[invalid] style_emoticons/default/smile.gif)

I don't use PS, but would a masking layer do it? Create a mask the shape of the bubble, now it will clip and still remain editable text.

I don't use PS, but would a masking layer do it? Create a mask the shape of the bubble, now it will clip and still remain editable text.

That would work too.

If I can be of help, I just took the separate letters, lined them up and group Free Transformed them to bend to my will, followed by individual touch ups.

Thank you all so much for the help (IMG:[invalid] style_emoticons/default/biggrin.gif)

Figured this would be a good place to ask. I'm trying to automate a couple of scan enhancements, e.g. i'm substituting colour leveling by unsharpen mask and automatically removing skewing based on an edge detector. But i have trouble with two things:

-------------
1. A decent standard for unsharp mask parameters. In some images i need to go as high as radius 10/gain 6, whilst in others i can do with radius 5/gain 0.5 . I lack the experience to tell me a "good enough" set of parameters.

(someone must be using unsharp masks as well as colour leveling here on EH, and certainly has more experience than i do)

-------------
2. How do you guys deal with the edge of the pages, like the left of this one. The best way i found was to layer out 1/4 of the left of the page (or the right if the bad edge is there) and colour level and unsharp heavily that layer. Then merge everything together.

The result is decent (note the left bottom corner):


But the process is hard to automate. I'd like to hear other ways that it can be done.

--------------------
IMPORTANT NOTE: i'm looking for a good enough not necessarily the best editing. I'll run this over 5000+ images, and if 90% percent of them come out alright that's good enough.

This post has been edited by blue penguin: Sep 15 2015, 20:21

I've thought a bit about this myself, and in fact a previous job had me involved in correcting scans of documents taken with cellphones. I'm not means an expert, but here are my thoughts.

0) Edge detection is simple enough, but what's hard is picking the ones that should be straight. Some pages don't have a straight edge (they are all slanted lines). Some pages don't have any clean lines. I think a better strategy would be text-based deskewing. Draw a box around each character, draw a line through the center of each one, take the angle, that's how much you need to deskew. More pages have clean text than have perfectly straight lines, so I think this would be a better approach (less mistakes / manual intervention needed).

2) I myself scan after debinding the book, so I don't often have to deal with curled edges. I found some really interesting literature about it though, with some very promising results. Here are some papers I think may set you on the right path (especially the first one):
[ieeexplore.ieee.org] http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1561180
[ieeexplore.ieee.org] http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1323798
[ieeexplore.ieee.org] http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1334172
[ieeexplore.ieee.org] http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1315007
[ieeexplore.ieee.org] http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1575694.

Some of those are very similar works by the same people. If you can't access IEEE, they may be hosted elsewhere or I can download them and send them to you.

This post has been edited by ilwaz: Sep 15 2015, 23:31

0) Edge detection is simple enough, but what's hard is picking the ones that should be straight. Some pages don't have a straight edge (they are all slanted lines). Some pages don't have any clean lines. I think a better strategy would be text-based deskewing. Draw a box around each character, draw a line through the center of each one, take the angle, that's how much you need to deskew. More pages have clean text than have perfectly straight lines, so I think this would be a better approach (less mistakes / manual intervention needed).

Interesting. What i have been doing 'till now is that i eliminated all not straight edges (with a threshold), then eliminated all edges that touched the border of the page, and then made a rectangle around (touching) the outer side of all edges. The smaller rectangle i could get (from several with rotations -10 to 10) assumed to be the orientation of the page.

The idea of text is interesting, but i'm not sure how to deal with japanese text (i'm a little clueless there). And if you allow me to add some criticism to the text idea, iv'e found a weakness in it after thinking a little: If you're transforming scanlated pages (i.e. already translated pages) there's a good change that the editor wrote horizontal text on a skewed page.

I know that transforming already scanlated stuff is bad practice, but hey, i'm trying to make a more-or-less idiot proof algorithm.

2) I myself scan after debinding the book, so I don't often have to deal with curled edges. I found some really interesting literature about it though, with some very promising results. Here are some papers I think may set you on the right path (especially the first one):
[ieeexplore.ieee.org] http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1561180
[ieeexplore.ieee.org] http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1323798
[ieeexplore.ieee.org] http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1334172
[ieeexplore.ieee.org] http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1315007
[ieeexplore.ieee.org] http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1575694.

Some of those are very similar works by the same people. If you can't access IEEE, they may be hosted elsewhere or I can download them and send them to you.

Wow, Mr. Zheng Zhang from NUS did a good deal of work on that. Thanks. And do not worry about IEEExplore, i use it almost daily.

Interesting. What i have been doing 'till now is that i eliminated all not straight edges (with a threshold), then eliminated all edges that touched the border of the page, and then made a rectangle around (touching) the outer side of all edges. The smaller rectangle i could get (from several with rotations -10 to 10) assumed to be the orientation of the page.

The idea of text is interesting, but i'm not sure how to deal with japanese text (i'm a little clueless there). And if you allow me to add some criticism to the text idea, iv'e found a weakness in it after thinking a little: If you're transforming scanlated pages (i.e. already translated pages) there's a good change that the editor wrote horizontal text on a skewed page.

I know that transforming already scanlated stuff is bad practice, but hey, i'm trying to make a more-or-less idiot proof algorithm.

Oh, interesting. So this'll be applied to already translated works as well. Then I have an unproven theory that you may want to test out. To supplement or augment your line edge detection, it seems to me that all homogeneous patterns in doujins (generally hair colors or areas of constant shading) are aligned to what should be level. Generally you can draw a line across a "row" of one of these patterns and discern the skew.

The logic for detecting this isn't nearly as clean as line edge detection, but I think there will be less false positives and more overall matches as some images will not have straight lines.

Then I have an unproven theory that you may want to test out. To supplement or augment your line edge detection, it seems to me that all homogeneous patterns in doujins (generally hair colors or areas of constant shading)

True, those things are computer generated these days and therefore always using the same positioning.

e.g. in the page above the pattern of the trap's leg shading is horizontal.

I see issues with small images (e.g. 600x1000 pixels) and with doujins from the '90s (i think they used transparent foil with the pattern during the '80s/'90s so it might be glued in a non-paraxial manner). Then again, the idea is worth trying just because it seems fun to make the experiment.

So guys, I been having some fun with this raws

[drive.google.com] https://drive.google.com/folderview?id=0Bwy...amp;usp=sharing

Cleaning samples are 3840x5850 keeping the printed aspect ratio, but that's just a random number at this point.





Looks nice at sight but no idea how they fare under technical scrutiny ? (IMG:[invalid] style_emoticons/default/unsure.gif)

One of first questions in my mind is how to calculate the optimal resolution and dpi to keep the best quality? tank being a common B5 size, and scans 7656x10524 900dpi

So guys, I been having some fun with this raws

[drive.google.com] https://drive.google.com/folderview?id=0Bwy...amp;usp=sharing

Cleaning samples are 3840x5850 keeping the printed aspect ratio, but that's just a random number at this point.





Looks nice at sight but no idea how they fare under technical scrutiny ? (IMG:[invalid] style_emoticons/default/unsure.gif)

One of first questions in my mind is how to calculate the optimal resolution and dpi to keep the best quality? tank being a common B5 size, and scans 7656x10524 900dpi

600dpi is usually the optimal size if you want "big" scans, which I readjust to 4200x6000. For the mass audience, I usually settle on half of that, or 2100x3000.

600dpi is usually the optimal size if you want "big" scans, which I readjust to 4200x6000. For the mass audience, I usually settle on half of that, or 2100x3000.

So 600dpi for optimal quality? Will remember that, ty.

But should I resize my PSDs to 4200x6000 at the very beginning or keep them original res all the time and only export different sizes ?

I heard that cleaning stuff on resolution higher than final target creates noise and black spots when downscaling, forcing you to clean twice.

This post has been edited by xanagu: Sep 19 2015, 10:25

So 600dpi for optimal quality? Will remember that, ty.

But should I resize my PSDs to 4200x6000 at the very beginning or keep them original res all the time and only export different sizes ?

I heard that cleaning stuff on resolution higher than final target creates noise and black spots when downscaling, forcing you to clean twice.

Always keep an unaltered back up set in case anything buggy happens. Wash, crop, rotate your raws to how you like them before downsizing, and I usually save them as PNG's. As for black spots, I've never heard of that one.

Sure a backup of the uncropped/unrotated raws is saved somewhere else;

but the PSD will stay always cropped resolution then (average 6420x9150) ?

Only exporting each page to different resolutions once everything is done, right?

As for black spots, I've never heard of that one.

Is not written on stone of course; I just saw it here

[youtu.be] https://youtu.be/ZVWt0hgxrAE?t=13m48s

and my raws are particulary big-ass sized, that's why (IMG:[invalid] style_emoticons/default/ohmy.gif)

This post has been edited by xanagu: Sep 19 2015, 12:53

I heard that cleaning stuff on resolution higher than final target creates noise and black spots when downscaling, forcing you to clean twice.

That's new to me. I tried to reproduce that issue in GIMP by doing what he did in the video (using another image of course), but with no luck. Makes me wonder if it's software, driver or algorithm related.

This post has been edited by Izur: Sep 19 2015, 14:02

Does anyone have the international version of Wildwords? Or any version of Wildwords Lower?

Does anyone have the international version of Wildwords? Or any version of Wildwords Lower?

 WWLower.zip ( 94.74k ) Number of downloads: 52

 WWLower.zip ( 94.74k ) Number of downloads: 52

Thanks! Hope another lovely person has CC Wildwords Int.

Hi there guys...!

So, a long time ago I messed around with PS but it never amounted to anything substantial.
This time however, I've spent time reading stuff and doing some experiments.
I've even taken up a bounty to ensure that I finish it. Here are the RAWs: http://ehentaihip.com/g/634337/2b0ba66b46/

And these are my workings.

1. [i.imgur.com] http://i.imgur.com/i9tWZVh.png
2. [i.imgur.com] http://i.imgur.com/Urz2gS2.png
3. [i.imgur.com] http://i.imgur.com/UjJjemO.png
4. [i.imgur.com] http://i.imgur.com/AF6K0mK.png
5. [i.imgur.com] http://i.imgur.com/6wWn5tE.png
6. [i.imgur.com] http://i.imgur.com/rO6QZQT.png
7. [i.imgur.com] http://i.imgur.com/mi0AJge.png
8. [i.imgur.com] http://i.imgur.com/NQSgIYD.png

I'm horrible at typesetting (specially SFX) and cloning is a fucking nightmare. Pages 4. and 6. just seem wrong but I'm quite proud of 8.

Anyone has tips?