HOCR output always sets textangle 180 and omits baseline info if Tesseract is compiled with --disable-legacy

[robertknight]

Basic Information

tesseract 5.3.0-19-ga3b9ac, compiled with --disable-legacy

Operating System

macOS 13 Ventura

Compiler

clang 14.0

Current Behavior

When Tesseract is compiled with --disable-legacy, hOCR output reports each line as being upside-down (textangle 180) and omits baseline information.

Steps to reproduce:

./configure --disable-legacy
./tesseract some-image.jpg output hocr

In the generated output.hocr file, ocr_line entries look like this:

<span class='ocr_line' id='line_1_142' title="bbox 1334 3054 2119 3088; textangle 180; x_size 34; x_descenders 8; x_ascenders 9">

Expected Behavior

If orientation information isn't available I'd expect the image to always be treated as if it were page-up. So entries should look like this:

<span class='ocr_line' id='line_1_142' title="bbox 1334 3054 2119 3088; baseline 0 -8; x_size 34; x_descenders 8; x_ascenders 9">

Suggested Fix

Internally, it looks like the issue is that:

1. ColumnFinder::text_rotation_ is initialized to a null vector. When the legacy engine is disabled, the ColumnFinder::CorrectOrientation function does not get called, and so this vector remains null.
2. This null vector gets propagated to PageIterator::Orientation, which does not handle this case correctly, as it converts this null vector to ORIENTATION_PAGE_DOWN -

   tesseract/src/ccmain/pageiterator.cpp

   Line 585 in a3b9acf

   if (up_in_image.y() > 0.0F) {

3. The HOCR renderer then maps this orientation value to textangle 180 and omits baseline info

Some fixes I tested locally were to change the initialization of ColumnFinder::text_rotation_ to be the same as the norotation value in ColumnFinder::CorrectOrientation, or to change the logic in PageIterator::Orientation to handle null rotation vectors by mapping them to ORIENTATION_PAGE_UP. I'm happy to submit a PR but I'm not sure the preferred way to go.

[amitdo]

#3997 seems related to this issue.

[amitdo]

When the legacy engine is disabled, the ColumnFinder::CorrectOrientation function does not get called

tesseract/src/ccmain/pagesegmain.cpp

Lines 335 to 406 in 67841aa

	#ifndef DISABLED_LEGACY_ENGINE
	if (equ_detect_) {
	equ_detect_->LabelSpecialText(to_block);
	}
	BLOBNBOX_CLIST osd_blobs;
	// osd_orientation is the number of 90 degree rotations to make the
	// characters upright. (See tesseract/osdetect.h for precise definition.)
	// We want the text lines horizontal, (vertical text indicates vertical
	// textlines) which may conflict (eg vertically written CJK).
	int osd_orientation = 0;
	bool vertical_text =
	textord_tabfind_force_vertical_text || pageseg_mode == PSM_SINGLE_BLOCK_VERT_TEXT;
	if (!vertical_text && textord_tabfind_vertical_text && PSM_ORIENTATION_ENABLED(pageseg_mode)) {
	vertical_text = finder->IsVerticallyAlignedText(textord_tabfind_vertical_text_ratio, to_block,
	&osd_blobs);
	}
	if (PSM_OSD_ENABLED(pageseg_mode) && osd_tess != nullptr && osr != nullptr) {
	std::vector<int> osd_scripts;
	if (osd_tess != this) {
	// We are running osd as part of layout analysis, so constrain the
	// scripts to those allowed by *this.
	AddAllScriptsConverted(unicharset, osd_tess->unicharset, &osd_scripts);
	for (auto &lang : sub_langs_) {
	AddAllScriptsConverted(lang->unicharset, osd_tess->unicharset, &osd_scripts);
	}
	}
	os_detect_blobs(&osd_scripts, &osd_blobs, osr, osd_tess);
	if (pageseg_mode == PSM_OSD_ONLY) {
	delete finder;
	return nullptr;
	}
	osd_orientation = osr->best_result.orientation_id;
	double osd_score = osr->orientations[osd_orientation];
	double osd_margin = min_orientation_margin * 2;
	for (int i = 0; i < 4; ++i) {
	if (i != osd_orientation && osd_score - osr->orientations[i] < osd_margin) {
	osd_margin = osd_score - osr->orientations[i];
	}
	}
	int best_script_id = osr->best_result.script_id;
	const char *best_script_str = osd_tess->unicharset.get_script_from_script_id(best_script_id);
	bool cjk = best_script_id == osd_tess->unicharset.han_sid() ||
	best_script_id == osd_tess->unicharset.hiragana_sid() ||
	best_script_id == osd_tess->unicharset.katakana_sid() ||
	strcmp("Japanese", best_script_str) == 0 ||
	strcmp("Korean", best_script_str) == 0 || strcmp("Hangul", best_script_str) == 0;
	if (cjk) {
	finder->set_cjk_script(true);
	}
	if (osd_margin < min_orientation_margin) {
	// The margin is weak.
	if (!cjk && !vertical_text && osd_orientation == 2) {
	// upside down latin text is improbable with such a weak margin.
	tprintf(
	"OSD: Weak margin (%.2f), horiz textlines, not CJK: "
	"Don't rotate.\n",
	osd_margin);
	osd_orientation = 0;
	} else {
	tprintf(
	"OSD: Weak margin (%.2f) for %d blob text block, "
	"but using orientation anyway: %d\n",
	osd_margin, osd_blobs.length(), osd_orientation);
	}
	}
	}
	osd_blobs.shallow_clear();
	finder->CorrectOrientation(to_block, vertical_text, osd_orientation);
	#endif // ndef DISABLED_LEGACY_ENGINE

I think we can fix the issue by enabling some parts of the code in this block instead of disabling the whole block of code when the legacy engine is disabled.