80 lines
3.4 KiB
Python
80 lines
3.4 KiB
Python
from ingestion.extract.glyph_order import find_reading_order_issues, is_reversed_order
|
|
|
|
|
|
def test_normal_ltr_span_not_flagged():
|
|
# ordinary increasing x-origins, as any normal left-to-right span has
|
|
assert not is_reversed_order([264.7, 269.4, 271.6, 276.3, 278.5])
|
|
|
|
|
|
def test_confirmed_page_1373_defect_shape_is_flagged():
|
|
# exact x-origins read via get_text("rawdict") from physical page 1373's
|
|
# affected span (" tịx 4 =" reversed) — see docs/pdf-parsing-outlier-catalog.md item 9
|
|
x_origins = [66.32, 64.17, 61.53, 58.89, 54.14, 51.98, 47.23]
|
|
assert is_reversed_order(x_origins)
|
|
|
|
|
|
def test_single_char_span_not_flagged():
|
|
assert not is_reversed_order([100.0])
|
|
|
|
|
|
def test_empty_span_not_flagged():
|
|
assert not is_reversed_order([])
|
|
|
|
|
|
def test_tied_x_origins_not_flagged_as_reversed():
|
|
# equal x-origins (e.g. stacked/overlapping glyphs) are not "decreasing"
|
|
assert not is_reversed_order([100.0, 100.0, 100.0])
|
|
|
|
|
|
def test_correctly_ordered_row_not_flagged():
|
|
row = {(20, 550.9): [(518.0, "n"), (525.2, "h"), (532.6, "i"), (536.8, "e")]}
|
|
assert find_reading_order_issues(row) == []
|
|
|
|
|
|
def test_confirmed_page_714_row_misorder_is_flagged():
|
|
# reproduces the real page-714 finding: within one PyMuPDF block (20),
|
|
# 4 line fragments are emitted out of x-order ("quản ", " ộ", "đ tệih",
|
|
# "n " concatenated) that reconstruct correctly ("...nhiệt độ") when
|
|
# re-sorted by x-origin — see outlier catalog item 9.
|
|
row = {
|
|
(20, 550.9): [
|
|
(518.06, "n"), (525.20, " "),
|
|
(546.59, " "), (553.71, "ộ"),
|
|
(541.84, "đ"), (539.45, " "), (536.81, "t"), (532.59, "ệ"), (529.95, "i"), (525.20, "h"),
|
|
]
|
|
}
|
|
issues = find_reading_order_issues(row)
|
|
assert len(issues) == 1
|
|
assert issues[0].extracted_text != issues[0].corrected_text
|
|
|
|
|
|
def test_different_blocks_at_same_y_not_merged():
|
|
# regression test for a real false positive: two DIFFERENT paragraphs in
|
|
# different PyMuPDF blocks (a right-column paragraph starting at x=299.4
|
|
# and a left-column paragraph starting at x=35.4, page 1104) coincide at
|
|
# the same y — grouping by block index (not a hand-picked x-coordinate
|
|
# column boundary) is what keeps them from being merged into one "row".
|
|
# This is the caller's responsibility (scan_reading_order groups by real
|
|
# PyMuPDF block index); find_reading_order_issues just trusts its input
|
|
# is already correctly grouped, which these two dict entries demonstrate.
|
|
row_block_1 = {(1, 70.4): [(299.39, "m"), (306.78, "ô")]}
|
|
row_block_4 = {(4, 70.4): [(35.43, "d"), (40.18, "e")]}
|
|
assert find_reading_order_issues(row_block_1) == []
|
|
assert find_reading_order_issues(row_block_4) == []
|
|
|
|
|
|
def test_kerning_jitter_not_flagged_as_reading_order_defect():
|
|
# regression test for a real false positive found by running against the
|
|
# actual PDF: "mefloquin" ('l' at x=491.566, 'o' at x=491.471 — a
|
|
# 0.095pt kerning-driven dip) was previously "corrected" into the wrong
|
|
# word "mefolquin". A row-level check with no decrease tolerance treats
|
|
# ordinary kerning as a defect and corrupts already-correct text.
|
|
row = {
|
|
(5, 449.7): [
|
|
(474.865, "m"), (482.161, "e"), (486.284, "f"),
|
|
(491.566, "l"), (491.471, "o"), (496.126, "q"),
|
|
(500.781, "u"), (505.436, "i"), (507.982, "n"),
|
|
]
|
|
}
|
|
assert find_reading_order_issues(row) == []
|