Implement Text Extraction in PyMuPdf Fitz Layout Mode

[MarcoWel]

trafficstars

Thank you for this excellent muPdf wrapper!

One feature that muPdf does not implement natively is layout-preserving plain text extraction.

• XPdf / poppler's pdftotext offer a layout mode as standard: https://www.mankier.com/1/pdftotext
• Other wrappers like PyMuPdf add their own implementation. Their fitz module extracts text in layout mode by default: python -m fitz gettext input.pdf https://pymupdf.readthedocs.io/en/latest/module.html#text-extraction

This is how the PyMuPdf fitz module does it: https://github.com/pymupdf/PyMuPDF/blob/main/fitz/main.py#L577

When layout preservation is a must, there is currently no other way than invoking pdftotext from the go app or - even nastier - calling the fitz python module from go.

How hard would it be to add this to go-fitz as well?

[MarcoWel]

I just had a closer look at how to possibly implement a layout-preserving func Text() in go.

A good starting point is checking the native C implementation for fz_new_buffer_from_stext_page: https://github.com/ArtifexSoftware/mupdf/blob/master/source/fitz/util.c#L424

fz_buffer *
fz_new_buffer_from_stext_page(fz_context *ctx, fz_stext_page *page)
{
	fz_stext_block *block;
	fz_stext_line *line;
	fz_stext_char *ch;
	fz_buffer *buf;

	buf = fz_new_buffer(ctx, 256);
	fz_try(ctx)
	{
		for (block = page->first_block; block; block = block->next)
		{
			if (block->type == FZ_STEXT_BLOCK_TEXT)
			{
				for (line = block->u.t.first_line; line; line = line->next)
				{
					for (ch = line->first_char; ch; ch = ch->next)
						fz_append_rune(ctx, buf, ch->c);
					fz_append_byte(ctx, buf, '\n');
				}
				fz_append_byte(ctx, buf, '\n');
			}
		}
	}
	fz_catch(ctx)
	{
		fz_drop_buffer(ctx, buf);
		fz_rethrow(ctx);
	}

	return buf;
}

Now looking at the crucial structs: https://github.com/ArtifexSoftware/mupdf/blob/master/include/mupdf/fitz/structured-text.h#L159

/**
	A text block is a list of lines of text (typically a paragraph),
	or an image.
*/
struct fz_stext_block
{
	int type;
	fz_rect bbox;
	union {
		struct { fz_stext_line *first_line, *last_line; } t;
		struct { fz_matrix transform; fz_image *image; } i;
	} u;
	fz_stext_block *prev, *next;
};

/**
	A text line is a list of characters that share a common baseline.
*/
struct fz_stext_line
{
	int wmode; /* 0 for horizontal, 1 for vertical */
	fz_point dir; /* normalized direction of baseline */
	fz_rect bbox;
	fz_stext_char *first_char, *last_char;
	fz_stext_line *prev, *next;
};

/**
	A text char is a unicode character, the style in which is
	appears, and the point at which it is positioned.
*/
struct fz_stext_char
{
	int c;
	int color; /* sRGB hex color */
	fz_point origin;
	fz_quad quad;
	float size;
	fz_font *font;
	fz_stext_char *next;
};

Those are not present in the go-fitz library (yet). The auto-generated go structs don't do the trick:

type _Ctype_struct_fz_stext_block struct {
	_type	_Ctype_int
	bbox	_Ctype_struct___7
	_	[4]byte
	u	[32]byte
	prev	*_Ctype_struct_fz_stext_block
	next	*_Ctype_struct_fz_stext_block
}

type _Ctype_struct_fz_stext_line struct {
	wmode		_Ctype_int
	dir		_Ctype_struct___28
	bbox		_Ctype_struct___7
	first_char	*_Ctype_struct_fz_stext_char
	last_char	*_Ctype_struct_fz_stext_char
	prev		*_Ctype_struct_fz_stext_line
	next		*_Ctype_struct_fz_stext_line
}

type _Ctype_struct_fz_stext_char struct {
	c	_Ctype_int
	color	_Ctype_int
	origin	_Ctype_struct___28
	quad	_Ctype_struct___29
	size	_Ctype_float
	font	*_Ctype_struct_fz_font
	next	*_Ctype_struct_fz_stext_char
}

First step would be to include proper definitions for those structs within go-fitz. Any help is appreciated!

[MarcoWel]

Okay, got the start right...

Structs:

type fzRect struct {
	X0, Y0 float32
	X1, Y1 float32
}

type fzPoint struct {
	X, Y float32
}

type fzQuad struct {
	Ul fzPoint
	Ur fzPoint
	Ll fzPoint
	Lr fzPoint
}

const (
	FZ_STEXT_BLOCK_TEXT  = 0
	FZ_STEXT_BLOCK_IMAGE = 1
)

type fzStextBlock struct {
	Type int32
	Bbox fzRect
	U    struct {
		T struct {
			FirstLine *fzStextLine
			LastLine  *fzStextLine
			_         [16]byte
		}
		// I struct {
		// 	Transform fzMatrix
		// 	Image     *fzImage
		// }
	}
	Prev *fzStextBlock
	Next *fzStextBlock
}

type fzStextLine struct {
	Wmode     int32
	Dir       fzPoint
	Bbox      fzRect
	FirstChar *fzStextChar
	LastChar  *fzStextChar
	Prev      *fzStextLine
	Next      *fzStextLine
}

type fzStextChar struct {
	C      int32
	Color  int32
	Origin fzPoint
	Quad   fzQuad
	Size   float32
	Font   unsafe.Pointer
	Next   *fzStextChar
}

Now the call to fz_new_buffer_from_stext_page() from go-fitz Text() can simply be replaced by a go port of the original function:

func (f *Document) Text(pageNumber int) (string, error) {
	...

	// buf := C.fz_new_buffer_from_stext_page(f.ctx, text)
	// defer C.fz_drop_buffer(f.ctx, buf)
	// str := C.GoString(C.fz_string_from_buffer(f.ctx, buf))

	str := ""
	block := (*fzStextBlock)(unsafe.Pointer(text.first_block))
	for block != nil {
		if block.Type == FZ_STEXT_BLOCK_TEXT {
			line := block.U.T.FirstLine
			for line != nil {
				char := line.FirstChar
				for char != nil {
					str += string(rune(char.C))
					char = char.Next
				}
				str += "\n"
				line = line.Next
			}
			str += "\n"
		}
		block = block.Next
	}
	return str, nil
}

We can go from here! :)

[gen2brain]

@MarcoWel If you or someone else manage to implement this I am willing to merge it. I don't have a plan or time to work on this.

[MarcoWel]

@gen2brain On it...