freelawproject
Bad Cross Hatched Redaction fails
X‑ray appears to only recognize redactions when the overlapping area is a solid color—specifically, solid black, white, or (I guess) red. I noticed this over the weekend while inspecting redacted PDFs. The following screenshot shows an example:
This behavior occurs because our current system uses a strict unicolor test (e.g. via pixmap.is_unicolor) to vet whether an area is redacted. In our tests, for example, bad_cross_hatched_redactions.pdf correctly recognizes 639 characters as being obscured, yet when we run a full redaction inspection via xray.inspect, no redactions are flagged. Cross-hatch patterns do not pass the strict unicolor test—even though they are clearly redacted—leading to false negatives.
Proposed Improvement:
Instead of relying solely on a strict unicolor check, we should introduce adaptive thresholding for our pixmaps. The idea is to process the pixmap for a given redaction area to convert it into a binary (black/white) image. By doing so, we would detect redactions that use cross-hatch patterns by converting them into a uniform color (e.g., all black). At the same time, legitimate text (like white text on a black background) should remain unaffected.
The goal is to quickly assess if the region contains a mix of colors that would indicate it is not uniformly redacted. Since performance is a primary concern, we could optimize further with NumPy in the future, but for now, keeping the dependencies unchanged, I propose a function that checks the brightness of pixels and stops scanning as soon as it finds both black and white pixels.
def multi_colored(pixmap, threshold=128) -> bool:
"""
Check if a pixmap contains both black and white pixels.
We interpret a pixel as "dark" if its brightness (the average of R, G, B)
is below the given threshold, and "light" otherwise. This function iterates
over the pixel data and stops as soon as it finds a pixel whose classification
differs from the first pixel.
:param pixmap: The fitz.Pixmap to check.
:param threshold: The brightness threshold (0-255) for classifying pixels.
:return: True if the pixmap contains both dark and light pixels, False otherwise.
"""
data = pixmap.samples
channels = pixmap.n
# Determine the classification for the first pixel.
first_brightness = (data[0] + data[1] + data[2]) // 3
first_is_dark = first_brightness < threshold
# Iterate over subsequent pixels.
for i in range(channels, len(data), channels):
brightness = (data[i] + data[i + 1] + data[i + 2]) // 3
is_dark = brightness < threshold
if is_dark != first_is_dark:
return True # Mixed colors detected.
return False
Then all we need to do is switch out
if not pixmap.is_unicolor:
for
if multi_colored(pixmap, threshold=128):
This method quickly determines if the redaction area is likely (aka text) or a solid block including the cross-hatch patterns.
Performance: By stopping as soon as both dark and light pixels are detected, we reduce unnecessary processing. I think it shouldn't be tremendously slower.
Accuracy: This approach should bridge the gap between our current system and real-world redaction implementations, potentially allowing us to correctly flag more redactions in our database.
Perhaps we could find a good number more redactions.
@mlissner