(muPDF) Forced OCR: wrong coordinates

[mergen3107]

• KOReader version: v2023.10-55
• Device: Kindle Scribe (5.16.2)

Issue

Enabling Forced OCR results in very wrong coordinates when long tapping to highlight text. Supporting my observations in https://github.com/koreader/koreader/issues/8068#issuecomment-1868435205

Steps to reproduce

Attached are an example file and a crash log with these steps: 1) Reopen KOReader with this file as last document, on page 1. 2) Tap and hold the whole Abstract, from the word Abstract. to purposes. in the very end. 3) It selected OK, select Highlight, then tap it again and delete. 4) Go to bottom menu, enable Forced OCR, tap outside to close bottom menu. 5) Try to highlight the same words in the Abstract. 6) Now it starts the highlight way lower, at the word Advanced - this is the first sentence in the 1. Introduction section below. 7) When I finish the highlight, it ends up highlighting Advanced Composition, then when I hit Highlight, it ultimately highlights Advanced Composition Explorer.

Highlights in the crash log are done by the SCribe pen, but it is same results with the finger as well. I use pen to be more precise when selecting words.

What other logs can I provide?

crash.log and the file.

exampleFile_OCR_wrong_coordinates.log.zip

crash_OCR_wrong_coordinates.log.txt

[mergen3107]

7. When I finish the highlight...

OK, weird stuff! If I continue from where I left off: 8) Now that the words Advanced Composition Explorer are selected, I go to the bottom menu and disable the Forced OCR. 9) The highlight of the words Advanced Composition Explorer disappears, and is replaced by a good highlight of the the whole abstract, as if it was done in with Forced OCR turned off.

So this enabling Forced OCR only results in showing the highlight wrong. Weird!

Here is rest of the log with steps 8 and beyond: crash_OCR_part2.log.txt

I played around the highlight again, don't remember the order.

[mergen3107]

At step 7, when only Advanced Composition Explorer is shown, going to the Bookmarks reveals that it is indeed the whole abstract that has been stored as the highlight. So, only the visual aspect of it is wrong.

Let me test the dictionary query...

[mergen3107]

OK, dictionary queries are affected. So, when Forced OCR is disabled, when I select Abstract. I get the results for the word abstract. When Forced OCR is enabled, the dictionary searches for Advancedcomposition without a whitespace.

Log: (strangely, crash.log did not update after these actions, so I can't provide it).

[mergen3107]

OCR files are the same as from here: https://github.com/koreader/koreader/issues/7860#issuecomment-862995754 (There are also the same ones I used to write up this Wiki page: https://github.com/koreader/koreader/wiki/Dictionary-support#dictionary-lookups-in-scanned-pages)

[mergen3107]

OCR itself seems to work great. From frontend/document/koptinterface.lua (around L520):

function KoptInterface:getTextBoxes(doc, pageno)
    local text = doc:getPageTextBoxes(pageno)
    logger.dbg("LOGG text =", text)
    if text and #text > 1 and doc.configurable.forced_ocr ~= 1 then
        return text
    -- if we have no text in original page then we will reuse native word boxes
    -- in reflow mode and find text boxes from scratch in non-reflow mode
    else
        if doc.configurable.text_wrap == 1 then
            return self:getNativeTextBoxes(doc, pageno)
        else
            return self:getNativeTextBoxesFromScratch(doc, pageno)
        end
    end
end

LOGG text = shows:

...
12/28/23-21:46:35 DEBUG LOGG text = {
  {
    {
      word = "SOLAR",
      x0 = 128.75999450683594,
      x1 = 166.51777648925781,
      y0 = 95.568199157714844,
      y1 = 110.65730285644531
    } --[[table: 0x7f16dc0fb078]],
    {
      word = "WIND",
      x0 = 168.93115234375,
      x1 = 199.876953125,
      y0 = 95.568199157714844,
      y1 = 110.65730285644531
    } --[[table: 0x7f16dc0fb598]],
    {
      word = "ELECTRON",
      x0 = 202.65058898925781,
      x1 = 263.80853271484375,
      y0 = 95.568199157714844,
      y1 = 110.65730285644531
    } --[[table: 0x7f16dc0fbd98]],
    {
      word = "PROTON",
      x0 = 266.221923828125,
      x1 = 312.03457641601563,
      y0 = 95.568199157714844,
      y1 = 110.65730285644531
    } --[[table: 0x7f16dc0fc430]],
...

So the culprit must be readerhighlight.lua then.

[mergen3107]

I am trying to trace the function involved in highlighting text while forced_ocr = 1. 1) I found with ripgrep that forced_ocr is meaningfully only present here: https://github.com/koreader/koreader/blob/63329569eb8afb00a29a723f215c6c5fcc9041aa/frontend/document/koptinterface.lua#L520-L533

KoptInterface:getTextBoxes(doc, pageno) is responsible for getting text boxes of the document. Depending on reflow, it then goes to function getNativeTextBoxes(doc, pageno) or getNativeTextBoxesFromScratch(doc, pageno).

The latter seems to work OK (tested with reflow off for now), as shown in the comment above.

2) When stuff is highlighted, I guess it starts here is the onHold function: https://github.com/koreader/koreader/blob/63329569eb8afb00a29a723f215c6c5fcc9041aa/frontend/apps/reader/modules/readerhighlight.lua#L1168-L1176

3) However, I cannot find the reference to KoptInterface:getTextBoxes(doc, pageno) in that function from readerhighlight.lua. I can only see it in function ReaderHighlight:onTranslateCurrentPage() and function ReaderHighlight:getExtendedHighlightPage().

Am I looking for the wrong function getTextBoxes()?

[mergen3107]

Am I looking for the wrong function getTextBoxes()?

Yes, this was wrong. What I need is this: https://github.com/koreader/koreader/blob/63329569eb8afb00a29a723f215c6c5fcc9041aa/frontend/apps/reader/modules/readerhighlight.lua#L1208

Now, onto getWordFromPosition...

[mergen3107]

I think I found the problem.

Here is the first two items in the output of getTextBoxes. Each item is duplicated. Is this how it supposed to be?

12/29/23-01:00:41 DEBUG LOGG-5 KI.lua getTextBoxes = {
  {
    {
      x0 = 312,
      x1 = 402,
      y0 = 138,
      y1 = 148
    } --[[table: 0x7feb9acc23c8]],
    x0 = 312,
    x1 = 402,
    y0 = 138,
    y1 = 148
  } --[[table: 0x7feb9acc2380]],
  {
    {
      x0 = 194,
      x1 = 220,
      y0 = 164,
      y1 = 174
    } --[[table: 0x7feb9ab6f008]],
    x0 = 194,
    x1 = 220,
    y0 = 164,
    y1 = 174
  } --[[table: 0x7feb93f8c108]],
...

This is the output at L530 here: https://github.com/koreader/koreader/blob/63329569eb8afb00a29a723f215c6c5fcc9041aa/frontend/document/koptinterface.lua#L520-L533

[mergen3107]

Nevermind, text from L521 is the same in both OCR on and OCR off cases, so that's not it.

[mergen3107]

Oh, wait, no, they are not the same!

Here is the no-OCR log of text - this is when text layers from PDF itself are taken.

12/29/23-01:00:41 DEBUG LOGG-5 text = {
  {
    {
      word = "564",
      x0 = 119.16000366210938,
      x1 = 135.43389892578125,
      y0 = 72.549667358398438,
      y1 = 87.463569641113281
    } --[[table: 0x7feb99577d20]],
    x0 = 119.16000366210938,
    x1 = 135.43389892578125,
    y0 = 72.549667358398438,
    y1 = 87.463569641113281
  } --[[table: 0x7feb9aad8a28]],
  {
    {
      word = "D.",
      x0 = 257.39999389648438,
      x1 = 265.15200805664063,
      y0 = 75.656021118164063,
      y1 = 86.552017211914063
    } --[[table: 0x7feb9ab65c08]],
    {
      word = "J.",
      x0 = 267.11203002929688,
      x1 = 272.20004272460938,
      y0 = 75.656021118164063,
      y1 = 86.552017211914063
    } --[[table: 0x7feb93f78e50]],
    {
      word = "McCOMAS",
      x0 = 274.16006469726563,
      x1 = 313.12814331054688,
      y0 = 75.656021118164063,
      y1 = 86.552017211914063
    } --[[table: 0x7feb9a95ed90]],
    {
      word = "ET",
      x0 = 315.2801513671875,
      x1 = 325.08816528320313,
      y0 = 75.656021118164063,
      y1 = 86.552017211914063
    } --[[table: 0x7feb9aaf1680]],
...

[mergen3107]

So, in the OCR mode, word = is missing completely (*from the tables), because it is nil (?)

[mergen3107]

I think I traced down where words stop appearing: https://github.com/koreader/koreader/blob/63329569eb8afb00a29a723f215c6c5fcc9041aa/frontend/document/koptinterface.lua#L674

The output of this function is the first one where there are no word = in the table:

12/29/23-02:13:08 DEBUG LOGG-6 getNativeTextBoxesFromScratch: RETURN boxes = {
  {
    {
      x0 = 312,
      x1 = 402,
      y0 = 138,
      y1 = 148
    } --[[table: 0x7f2328f00ed8]],
    x0 = 312,
    x1 = 402,
    y0 = 138,
    y1 = 148
  } --[[table: 0x7f2328f00e90]],
  {
    {
      x0 = 194,
      x1 = 220,
      y0 = 164,
      y1 = 174
    } --[[table: 0x7f2328f01038]],
    x0 = 194,
    x1 = 220,
    y0 = 164,
    y1 = 174
  } --[[table: 0x7f2328f00ff0]],
...

[mergen3107]

Then getNativeWordBoxes is from koreader/base/ffi/koptcontext.lua:

function KOPTContext_mt.__index:getNativeWordBoxes(bmp, x, y, w, h)
    return self:getWordBoxes(bmp, x, y, w, h, 1)
end

which uses this function above.

function KOPTContext_mt.__index:getWordBoxes(bmp, x, y, w, h, box_type)
    local boxa = ffi.new("BOXA[1]")
    local nai = ffi.new("NUMA[1]")
    local counter_l = ffi.new("int[1]")
    local nr_word, current_line
    local counter_w, counter_cw
    local l_x0, l_y0, l_x1, l_y1

    if box_type == 0 then
        k2pdfopt.k2pdfopt_get_reflowed_word_boxes(self,
            bmp == "src" and self.src or self.dst, x, y, w, h)
        boxa = self.rboxa
        nai = self.rnai
    elseif box_type == 1 then
        k2pdfopt.k2pdfopt_get_native_word_boxes(self,
            bmp == "src" and self.src or self.dst, x, y, w, h)
        boxa = self.nboxa
        nai = self.nnai
    end

    if boxa == nil or nai == nil then return end

    -- get number of lines in this area
    nr_word = leptonica.boxaGetCount(boxa)
    assert(nr_word == leptonica.numaGetCount(nai))

    local boxes = {}
    counter_w = 0
    while counter_w < nr_word do
        leptonica.numaGetIValue(nai, counter_w, counter_l)
        current_line = counter_l[0]
        -- sub-table that contains words in a line
        local lbox = {}
        boxes[counter_l[0]+1] = lbox
        counter_cw = 0
        l_x0, l_y0, l_x1, l_y1 = 9999, 9999, 0, 0
        while current_line == counter_l[0] and counter_w < nr_word do
            local box = leptonica.boxaGetBox(boxa, counter_w, C.L_CLONE)
            -- update line box
            l_x0 = box.x < l_x0 and box.x or l_x0
            l_y0 = box.y < l_y0 and box.y or l_y0
            l_x1 = box.x + box.w > l_x1 and box.x + box.w or l_x1
            l_y1 = box.y + box.h > l_y1 and box.y + box.h or l_y1
            -- box for a single word
            lbox[counter_cw+1] = {
                x0 = box.x, y0 = box.y,
                x1 = box.x + box.w,
                y1 = box.y + box.h,
            }
            counter_w, counter_cw = counter_w + 1, counter_cw + 1
            if counter_w < nr_word then
                leptonica.numaGetIValue(nai, counter_w, counter_l)
            end
        end
        if current_line ~= counter_l[0] then counter_w = counter_w - 1 end
        -- box for a whole line
        lbox.x0, lbox.y0, lbox.x1, lbox.y1 = l_x0, l_y0, l_x1, l_y1
        counter_w = counter_w + 1
    end
    return boxes, nr_word
end

I think the culprit is somewhere here :D

[mergen3107]

Why doesn't this OCR function isn't used anywhere? https://github.com/koreader/koreader/blob/63329569eb8afb00a29a723f215c6c5fcc9041aa/frontend/document/koptinterface.lua#L721-L730

[mergen3107]

I can only see it here, but this is not onHold function, but lookup: https://github.com/koreader/koreader/blob/63329569eb8afb00a29a723f215c6c5fcc9041aa/frontend/apps/reader/modules/readerhighlight.lua#L1437

[mergen3107]

Ah, OK. lookup( is called from function ReaderHighlight:onHoldRelease().

However, the coordinates of the boxes from the OP are wrong already at onHold.

[poire-z]

(Not following you on your investigation, all this code is quite obscure to many of us...) Just mentionning #3688 and that it may be related to DPI (joy!).

[mergen3107]

This function is really weird koptcontext.lua:

function KOPTContext_mt.__index:getWordBoxes(bmp, x, y, w, h, box_type)
    local boxa = ffi.new("BOXA[1]")
    local nai = ffi.new("NUMA[1]")
    local counter_l = ffi.new("int[1]")
    local nr_word, current_line
    local counter_w, counter_cw
    local l_x0, l_y0, l_x1, l_y1

    if box_type == 0 then
        k2pdfopt.k2pdfopt_get_reflowed_word_boxes(self,
            bmp == "src" and self.src or self.dst, x, y, w, h)
        boxa = self.rboxa
        nai = self.rnai
    elseif box_type == 1 then
        k2pdfopt.k2pdfopt_get_native_word_boxes(self,
            bmp == "src" and self.src or self.dst, x, y, w, h)
        boxa = self.nboxa
        nai = self.nnai
        logger.dbg("LOGGBASE boxa =", boxa)
        logger.dbg("LOGGBASE nai =", nai)
    end

    if boxa == nil or nai == nil then return end

    -- get number of lines in this area
    nr_word = leptonica.boxaGetCount(boxa)
    assert(nr_word == leptonica.numaGetCount(nai))

    local boxes = {}
    counter_w = 0
    while counter_w < nr_word do
        leptonica.numaGetIValue(nai, counter_w, counter_l)
        current_line = counter_l[0]
        -- sub-table that contains words in a line
        local lbox = {}
        boxes[counter_l[0]+1] = lbox
        counter_cw = 0
        l_x0, l_y0, l_x1, l_y1 = 9999, 9999, 0, 0
        while current_line == counter_l[0] and counter_w < nr_word do
            local box = leptonica.boxaGetBox(boxa, counter_w, C.L_CLONE)
            -- update line box
            l_x0 = box.x < l_x0 and box.x or l_x0
            l_y0 = box.y < l_y0 and box.y or l_y0
            l_x1 = box.x + box.w > l_x1 and box.x + box.w or l_x1
            l_y1 = box.y + box.h > l_y1 and box.y + box.h or l_y1
            -- box for a single word
            lbox[counter_cw+1] = {
                x0 = box.x, y0 = box.y,
                x1 = box.x + box.w,
                y1 = box.y + box.h,
            }
            counter_w, counter_cw = counter_w + 1, counter_cw + 1
            if counter_w < nr_word then
                leptonica.numaGetIValue(nai, counter_w, counter_l)
            end
        end
        if current_line ~= counter_l[0] then counter_w = counter_w - 1 end
        -- box for a whole line
        lbox.x0, lbox.y0, lbox.x1, lbox.y1 = l_x0, l_y0, l_x1, l_y1
        counter_w = counter_w + 1
    end
    return boxes, nr_word
end

It seems that boxes stay empty, because all involved variables are inside the while loop, which makes those vars empty at each cycle.

[mergen3107]

(Not following you on your investigation, all this code is quite obscure to many of us...) Just mentionning #3688 and that it may be related to DPI (joy!).

@poire-z Thanks! However, I really think the problem stems from the fact that word variables are not present when forced_ocr is on.

Coordinates seem sane. See example log from emulator:

12/29/23-02:43:53 DEBUG hold position in page {
  page = 2,
  rotation = 0,
  x = 183.32065597922661,
  y = 362.5031481665971,
  zoom = 1.2082653687705533
} --[[table: 0x7f1d7c14fa18]] 
12/29/23-02:43:53 DEBUG LOGG-5 text = {
  {
    {
      word = "564",
      x0 = 119.16000366210938,
      x1 = 135.43389892578125,
      y0 = 72.549667358398438,
      y1 = 87.463569641113281
    } --[[table: 0x7f1da10e2760]],
    x0 = 119.16000366210938,
    x1 = 135.43389892578125,
    y0 = 72.549667358398438,
    y1 = 87.463569641113281
  } --[[table: 0x7f1da1382ad0]],
  {
    {
      word = "D.",
      x0 = 257.39999389648438,
      x1 = 265.15200805664063,
      y0 = 75.656021118164063,
      y1 = 86.552017211914063
    } --[[table: 0x7f1da10d56d8]],

[mergen3107]

This function is really weird koptcontext.lua

Alright, I had some sleep :D this function is fine:

12/29/23-12:31:26 DEBUG LOGGBASE boxes END = {
  {
    {
      x0 = 120,
      x1 = 405,
      y0 = 177,
      y1 = 187
    } --[[table: 0x7f152ef18750]],
    x0 = 120,
    x1 = 405,
    y0 = 177,
    y1 = 187
  } --[[table: 0x7f152ef17830]],
  {
    {
      x0 = 356,
      x1 = 405,
      y0 = 229,
      y1 = 239
    } --[[table: 0x7f152cd3de50]],
    x0 = 356,
    x1 = 405,
    y0 = 229,
    y1 = 239
  } --[[table: 0x7f152cd3c578]]
} --[[table: 0x7f152ef17780]] 

It keeps populating the boxes table. However, it doesn't have word values in it, but the other function is waiting for them: https://github.com/koreader/koreader/blob/63329569eb8afb00a29a723f215c6c5fcc9041aa/frontend/document/koptinterface.lua#L901-L917

Here boxes are expected to have word, but the come from that function without them.

I'll write up the full chain now.

[mergen3107]

1) readerhighlight.lua, onHold: calls document-specific getWordFromPosition: https://github.com/koreader/koreader/blob/63329569eb8afb00a29a723f215c6c5fcc9041aa/frontend/apps/reader/modules/readerhighlight.lua#L1208 Input: document name/path doc, highlight holding position pos. Expected output: table with five values: word, pos0, pos1, sbox, pbox. The latter two are box on screen and box on page, respectively. sbox is for reflowed content, so in normal view sbox is nil.

2) The function getWordFromPosition from koptinterface.lua: https://github.com/koreader/koreader/blob/63329569eb8afb00a29a723f215c6c5fcc9041aa/frontend/document/koptinterface.lua#L1039-L1049 First, it needs local text_boxes = self:getTextBoxes(doc, pos.page) from itself: https://github.com/koreader/koreader/blob/63329569eb8afb00a29a723f215c6c5fcc9041aa/frontend/document/koptinterface.lua#L520-L533

Here we follow the last return self:getNativeTextBoxesFromScratch(doc, pageno) because forced_ocr=1 and there is no reflow. https://github.com/koreader/koreader/blob/63329569eb8afb00a29a723f215c6c5fcc9041aa/frontend/document/koptinterface.lua#L660-L684

Here we need local boxes, nr_word = kc:getNativeWordBoxes("src", 0, 0, page_size.w, page_size.h), which is in koreader-base/ffi/koptcontext.lua:

function KOPTContext_mt.__index:getNativeWordBoxes(bmp, x, y, w, h)
    return self:getWordBoxes(bmp, x, y, w, h, 1)
end

https://github.com/koreader/koreader-base/blob/d83148b27e25b9332678221f6737d66be7e73318/ffi/koptcontext.lua#L155-L157

which is then

function KOPTContext_mt.__index:getWordBoxes(bmp, x, y, w, h, box_type)
    local boxa = ffi.new("BOXA[1]")
    local nai = ffi.new("NUMA[1]")
    local counter_l = ffi.new("int[1]")
    local nr_word, current_line
    local counter_w, counter_cw
    local l_x0, l_y0, l_x1, l_y1

    if box_type == 0 then
        k2pdfopt.k2pdfopt_get_reflowed_word_boxes(self,
            bmp == "src" and self.src or self.dst, x, y, w, h)
        boxa = self.rboxa
        nai = self.rnai
    elseif box_type == 1 then
        k2pdfopt.k2pdfopt_get_native_word_boxes(self,
            bmp == "src" and self.src or self.dst, x, y, w, h)
        boxa = self.nboxa
        nai = self.nnai
    end

    if boxa == nil or nai == nil then return end

    -- get number of lines in this area
    nr_word = leptonica.boxaGetCount(boxa)
    assert(nr_word == leptonica.numaGetCount(nai))

    local boxes = {}
    counter_w = 0
    while counter_w < nr_word do
        leptonica.numaGetIValue(nai, counter_w, counter_l)
        current_line = counter_l[0]
        -- sub-table that contains words in a line
        local lbox = {}
        boxes[counter_l[0]+1] = lbox
        counter_cw = 0
        l_x0, l_y0, l_x1, l_y1 = 9999, 9999, 0, 0
        while current_line == counter_l[0] and counter_w < nr_word do
            local box = leptonica.boxaGetBox(boxa, counter_w, C.L_CLONE)
            -- update line box
            l_x0 = box.x < l_x0 and box.x or l_x0
            l_y0 = box.y < l_y0 and box.y or l_y0
            l_x1 = box.x + box.w > l_x1 and box.x + box.w or l_x1
            l_y1 = box.y + box.h > l_y1 and box.y + box.h or l_y1
            -- box for a single word
            lbox[counter_cw+1] = {
                x0 = box.x, y0 = box.y,
                x1 = box.x + box.w,
                y1 = box.y + box.h,
            }
            counter_w, counter_cw = counter_w + 1, counter_cw + 1
            if counter_w < nr_word then
                leptonica.numaGetIValue(nai, counter_w, counter_l)
            end
        end
        if current_line ~= counter_l[0] then counter_w = counter_w - 1 end
        -- box for a whole line
        lbox.x0, lbox.y0, lbox.x1, lbox.y1 = l_x0, l_y0, l_x1, l_y1
        counter_w = counter_w + 1
    end
    return boxes, nr_word
end

https://github.com/koreader/koreader-base/blob/d83148b27e25b9332678221f6737d66be7e73318/ffi/koptcontext.lua#L89-L149

The output of boxes here is the nested tables like:

12/29/23-12:31:26 DEBUG LOGGBASE boxes END = {
  {
    {
      x0 = 120,
      x1 = 405,
      y0 = 177,
      y1 = 187
    } --[[table: 0x7f152ef18750]],
    x0 = 120,
    x1 = 405,
    y0 = 177,
    y1 = 187
  } --[[table: 0x7f152ef17830]],
  {
    {
      x0 = 356,
      x1 = 405,
      y0 = 229,
      y1 = 239
    } --[[table: 0x7f152cd3de50]],
    x0 = 356,
    x1 = 405,
    y0 = 229,
    y1 = 239
  } --[[table: 0x7f152cd3c578]]
...

3) Now that the boxes are defined, the function getWordFromPosition from (2) calls getWordFromNativePosition: https://github.com/koreader/koreader/blob/63329569eb8afb00a29a723f215c6c5fcc9041aa/frontend/document/koptinterface.lua#L1089-L1098

This in turn needs getWordFromBoxes function: https://github.com/koreader/koreader/blob/63329569eb8afb00a29a723f215c6c5fcc9041aa/frontend/document/koptinterface.lua#L901-L917

So, here it expects that boxes have words as well, but they don't:

local wb = boxes[i][j]
...

        return {
            word = wb.word,
            box = box,
        }

Where should words come from in this chain?

[poire-z]

(Still not wanting to look at all this code :)). But I remember (crengine only, dunno) that there are legitimate cases where stuff like getWord fails (may be returning no "word"), and we fallback to getText which does the same thing (but differently, usually better) when the start and end pos are identical. So, be sure there's not something after this getWord stuff that makes it work using getText.

[mergen3107]

OK, but this happens not during onHold, but at highlightFromHoldPos: https://github.com/koreader/koreader/blob/63329569eb8afb00a29a723f215c6c5fcc9041aa/frontend/apps/reader/modules/readerhighlight.lua#L1672-L1686

and eventually getTextFromBoxes also assumes that boxes have words: https://github.com/koreader/koreader/blob/63329569eb8afb00a29a723f215c6c5fcc9041aa/frontend/document/koptinterface.lua#L941

[mergen3107]

From linked "PDF is too dark" issue:

By the way, I tried selecting words in the CBZ version with and without Forced OCR on, and ... it works flawless regardless O_O So my old issue with wrong coordinates when Forced OCR is on refers only to PDFs...