 |
Leptonica
1.82.0
Image processing and image analysis suite
|
|
Leptonica
1.82.0
Image processing and image analysis suite
|
#include <math.h>#include "allheaders.h"Go to the source code of this file.
Functions | |
| static void | pixDebugFlipDetect (const char *filename, PIX *pixs, PIX *pixhm, l_int32 enable) |
| PIX * | pixOrientCorrect (PIX *pixs, l_float32 minupconf, l_float32 minratio, l_float32 *pupconf, l_float32 *pleftconf, l_int32 *protation, l_int32 debug) |
| l_ok | pixOrientDetect (PIX *pixs, l_float32 *pupconf, l_float32 *pleftconf, l_int32 mincount, l_int32 debug) |
| l_ok | makeOrientDecision (l_float32 upconf, l_float32 leftconf, l_float32 minupconf, l_float32 minratio, l_int32 *porient, l_int32 debug) |
| l_ok | pixUpDownDetect (PIX *pixs, l_float32 *pconf, l_int32 mincount, l_int32 npixels, l_int32 debug) |
| l_ok | pixMirrorDetect (PIX *pixs, l_float32 *pconf, l_int32 mincount, l_int32 debug) |
High-level interface for detection and correction
PIX *pixOrientCorrect() Page orientation detection (pure rotation by 90 degree increments):
l_int32 pixOrientDetect()
l_int32 makeOrientDecision()
l_int32 pixUpDownDetect() Page mirror detection (flip 180 degrees about line in plane of image):
l_int32 pixMirrorDetect() Static debug helper
static void pixDebugFlipDetect()===================================================================
Page transformation detection:
Once a page is deskewed, there are 8 possible states that it can be in, shown symbolically below. Suppose state 0 is correct.
0: correct 1 2 3
+------+ +------+ +------+ +------+
| **** | | * | | **** | | * |
| * | | * | | * | | * |
| * | | **** | | * | | **** |
+------+ +------+ +------+ +------+ 4 5 6 7
+-----+ +-----+ +-----+ +-----+
| *** | | * | | *** | | * |
| * | | * | | * | | * |
| * | | * | | * | | * |
| * | | *** | | * | | *** |
+-----+ +-----+ +-----+ +-----+ Each of the other seven can be derived from state 0 by applying some
combination of a 90 degree clockwise rotation, a flip about
a horizontal line, and a flip about a vertical line,
all abbreviated as:
R = Rotation (about a line perpendicular to the image)
H = Horizontal flip (about a vertical line in the plane of the image)
V = Vertical flip (about a horizontal line in the plane of the image) We get these transformations:
RHV
000 -> 0
001 -> 1
010 -> 2
011 -> 3
100 -> 4
101 -> 5
110 -> 6
111 -> 7 Note that in four of these, the sum of H and V is 1 (odd).
For these four, we have a change in parity (handedness) of
the image, and the transformation cannot be performed by
rotation about a vertical line out of the page. Under
rotation R, the set of 8 transformations decomposes into
two subgroups linking {0, 3, 4, 7} and {1, 2, 5, 6} independently.pixOrientDetect() tests for a pure rotation (0, 90, 180, 270 degrees). It doesn't change parity.
pixMirrorDetect() tests for a horizontal flip about the vertical axis. It changes parity.
The landscape/portrait rotation can be detected in two ways:
(1) Compute the deskew confidence for an image segment,
both as is and rotated 90 degrees (see skew.c). (2) Compute the ascender/descender signal for the image,
both as is and rotated 90 degrees (implemented here). The ascender/descender signal is useful for determining text
orientation in Roman alphabets because the incidence of letters
with straight-line ascenders (b, d, h, k, l, 't') outnumber
those with descenders ('g', p, q). The letters 't' and 'g'
will respond variably to the filter, depending on the type face.What about the mirror image situations? These aren't common unless you're dealing with film, for example. But you can reliably test if the image has undergone a parity-changing flip once about some axis in the plane of the image, using pixMirrorDetect*(). This works ostensibly by counting the number of characters with ascenders that stick out to the left and right of the ascender. Characters that are not mirror flipped are more likely to extend to the right (b, h, k) than to the left (d). Of course, that is for text that is rightside-up. So before you apply the mirror test, it is necessary to insure that the text has the ascenders going up, and not down or to the left or right. But here's what *really* happens. It turns out that the pre-filtering before the hit-miss transform (HMT) is crucial, and surprisingly, when the pre-filtering is chosen to generate a large signal, the majority of the signal comes from open regions of common lower-case letters such as 'e', 'c' and 'f'.
The set of operations you actually use depends on your prior knowledge:
(1) If the page is known to be either rightside-up or upside-down, use
either pixOrientDetect() with pleftconf = NULL, or
pixUpDownDetect().(2) If any of the four orientations are possible, use pixOrientDetect().
(3) If the text is horizontal and rightside-up, the only remaining
degree of freedom is a left-right mirror flip: use pixMirrorDetect(). (4) If you have a relatively large amount of numbers on the page,
use the slower pixUpDownDetect().We summarize the full orientation and mirror flip detection process:
(1) First determine which of the four 90 degree rotations
causes the text to be rightside-up. This can be done
with either skew confidence or the pixOrientDetect()
signals. For the latter, see the table for pixOrientDetect().(2) Then, with ascenders pointing up, apply pixMirrorDetect(). In the normal situation the confidence confidence will be large and positive. However, if mirror flipped, the confidence will be large and negative.
A high-level interface, pixOrientCorrect() combines the detection of the orientation with the rotation decision and the rotation itself.
For pedagogical reasons, we have included a dwa implementation of this functionality, in flipdetectdwa.c.notused. It shows by example how to make a dwa implementation of an application that uses binary morphological operations. It is faster than the rasterop implementation, but not by a large amount.
Finally, use can be made of programs such as exiftool and convert to
read exif camera orientation data in jpeg files and conditionally rotate.
Here is an example shell script, made by Dan9er:
==================================================================
#!/bin/sh
# orientByExif.sh
# Dependencies: exiftool (exiflib) and convert (ImageMagick)
# Note: if there is no exif orientation data in the jpeg file,
# this simply copies the input file.
#
if [[ -z $(command -v exiftool) || -z $(command -v convert) ]]; then
echo "You need to install dependencies; e.g.:"
echo " sudo apt install libimage-exiftool-perl"
echo " sudo apt install imagemagick"
exit 1
fi
if [[ $# != 2 ]]; then
echo "Syntax: orientByExif infile outfile"
exit 2
fi
if [[ ${1: -4} != ".jpg" ]]; then
echo "File is not a jpeg"
exit 3
fi
if [[ $(exiftool -s3 -n -Orientation "$1") = 1 ]]; then
echo "Image is already upright"
exit 0
fi
convert "$1" -auto-orient "$2"
echo "Done"
exit 0
==================================================================
Definition in file flipdetect.c.
| l_ok makeOrientDecision | ( | l_float32 | upconf, |
| l_float32 | leftconf, | ||
| l_float32 | minupconf, | ||
| l_float32 | minratio, | ||
| l_int32 * | porient, | ||
| l_int32 | debug | ||
| ) |
| [in] | upconf | nonzero |
| [in] | leftconf | nonzero |
| [in] | minupconf | minimum value for which a decision can be made |
| [in] | minratio | minimum conf ratio required for a decision |
| [out] | porient | text orientation enum {0,1,2,3,4} |
| [in] | debug | 1 for debug output; 0 otherwise |
Notes:
(1) This can be run after pixOrientDetect()
(2) Both upconf and leftconf must be nonzero; otherwise the
orientation cannot be determined.
(3) The abs values of the input confidences are compared to
minupconf.
(4) The abs value of the largest of (upconf/leftconf) and
(leftconf/upconf) is compared with minratio.
(5) Input 0.0 for the default values for minupconf and minratio.
(6) The return value of orient is interpreted thus:
L_TEXT_ORIENT_UNKNOWN: not enough evidence to determine
L_TEXT_ORIENT_UP: text rightside-up
L_TEXT_ORIENT_LEFT: landscape, text up facing left
L_TEXT_ORIENT_DOWN: text upside-down
L_TEXT_ORIENT_RIGHT: landscape, text up facing right
Definition at line 464 of file flipdetect.c.
References L_TEXT_ORIENT_UNKNOWN.
| l_ok pixMirrorDetect | ( | PIX * | pixs, |
| l_float32 * | pconf, | ||
| l_int32 | mincount, | ||
| l_int32 | debug | ||
| ) |
| [in] | pixs | 1 bpp, deskewed, English text |
| [out] | pconf | confidence that text is not LR mirror reversed |
| [in] | mincount | min number of left + right; use 0 for default |
| [in] | debug | 1 for debug output; 0 otherwise |
Notes:
(1) For this test, it is necessary that the text is horizontally
oriented, with ascenders going up.
(2) conf is the normalized difference between the number of
right and left facing characters with ascenders.
Left-facing are {d}; right-facing are {b, h, k}.
At least that was the expectation. In practice, we can
really just say that it is the normalized difference in
hits using two specific hit-miss filters, textsel1 and textsel2,
after the image has been suitably pre-filtered so that
these filters are effective. See (4) for what's really happening.
(3) A large positive conf value indicates normal text, whereas
a large negative conf value means the page is mirror reversed.
(4) The implementation is a bit tricky. The general idea is
to fill the x-height part of characters, but not the space
between them, before doing the HMT. This is done by
finding pixels added using two different operations -- a
horizontal close and a vertical dilation -- and adding
the intersection of these sets to the original. It turns
out that the original intuition about the signal was largely
in error: much of the signal for right-facing characters
comes from the lower part of common x-height characters, like
the e and c, that remain open after these operations.
So it's important that the operations to close the x-height
parts of the characters are purposely weakened sufficiently
to allow these characters to remain open. The wonders
of morphology!
Definition at line 718 of file flipdetect.c.
| PIX* pixOrientCorrect | ( | PIX * | pixs, |
| l_float32 | minupconf, | ||
| l_float32 | minratio, | ||
| l_float32 * | pupconf, | ||
| l_float32 * | pleftconf, | ||
| l_int32 * | protation, | ||
| l_int32 | debug | ||
| ) |
| [in] | pixs | 1 bpp, deskewed, English text, 150 - 300 ppi |
| [in] | minupconf | minimum value for which a decision can be made |
| [in] | minratio | minimum conf ratio required for a decision |
| [out] | pupconf | [optional] ; use NULL to skip |
| [out] | pleftconf | [optional] ; use NULL to skip |
| [out] | protation | [optional] ; use NULL to skip |
| [in] | debug | 1 for debug output; 0 otherwise |
Notes:
(1) Simple top-level function to detect if Roman text is in
reading orientation, and to rotate the image accordingly if not.
(2) Returns a copy if no rotation is needed.
(3) See notes for pixOrientDetect() and pixOrientDecision().
Use 0.0 for default values for minupconf and minratio
(4) Optional output of intermediate confidence results and
the rotation performed on pixs.
Definition at line 274 of file flipdetect.c.
| l_ok pixOrientDetect | ( | PIX * | pixs, |
| l_float32 * | pupconf, | ||
| l_float32 * | pleftconf, | ||
| l_int32 | mincount, | ||
| l_int32 | debug | ||
| ) |
| [in] | pixs | 1 bpp, deskewed, English text, 150 - 300 ppi |
| [out] | pupconf | [optional] ; may be NULL |
| [out] | pleftconf | [optional] ; may be NULL |
| [in] | mincount | min number of up + down; use 0 for default |
| [in] | debug | 1 for debug output; 0 otherwise |
Notes:
(1) See "Measuring document image skew and orientation"
Dan S. Bloomberg, Gary E. Kopec and Lakshmi Dasari
IS&T/SPIE EI'95, Conference 2422: Document Recognition II
pp 302-316, Feb 6-7, 1995, San Jose, CA
(2) upconf is the normalized difference between up ascenders
and down ascenders. The image is analyzed without rotation
for being rightside-up or upside-down. Set &upconf to null
to skip this operation.
(3) leftconf is the normalized difference between up ascenders
and down ascenders in the image after it has been
rotated 90 degrees clockwise. With that rotation, ascenders
projecting to the left in the source image will project up
in the rotated image. We compute this by rotating 90 degrees
clockwise and testing for up and down ascenders. Set
&leftconf to null to skip this operation.
(4) Note that upconf and leftconf are not linear measures of
confidence, e.g., in a range between 0 and 100. They
measure how far you are out on the tail of a (presumably)
normal distribution. For example, a confidence of 10 means
that it is nearly certain that the difference did not
happen at random. However, these values must be interpreted
cautiously, taking into consideration the estimated prior
for a particular orientation or mirror flip. The up-down
signal is very strong if applied to text with ascenders
up and down, and relatively weak for text at 90 degrees,
but even at 90 degrees, the difference can look significant.
For example, suppose the ascenders are oriented horizontally,
but the test is done vertically. Then upconf can
be < -MIN_CONF_FOR_UP_DOWN, suggesting the text may be
upside-down. However, if instead the test were done
horizontally, leftconf will be very much larger
(in absolute value), giving the correct orientation.
(5) If you compute both upconf and leftconf, and there is
sufficient signal, the following table determines the
cw angle necessary to rotate pixs so that the text is
rightside-up:
0 deg : upconf >> 1, abs(upconf) >> abs(leftconf)
90 deg : leftconf >> 1, abs(leftconf) >> abs(upconf)
180 deg : upconf << -1, abs(upconf) >> abs(leftconf)
270 deg : leftconf << -1, abs(leftconf) >> abs(upconf)
(6) One should probably not interpret the direction unless
there are a sufficient number of counts for both orientations,
in which case neither upconf nor leftconf will be 0.0.
(7) This algorithm will fail on some images, such as tables,
where most of the characters are numbers and appear as
uppercase, but there are some repeated words that give a
biased signal. It may be advisable to run a table detector
first (e.g., pixDecideIfTable()), and not run the orientation
detector if it is a table.
(8) Uses rasterop implementation of HMT.
Definition at line 405 of file flipdetect.c.
| l_ok pixUpDownDetect | ( | PIX * | pixs, |
| l_float32 * | pconf, | ||
| l_int32 | mincount, | ||
| l_int32 | npixels, | ||
| l_int32 | debug | ||
| ) |
| [in] | pixs | 1 bpp, deskewed, English text, 150 - 300 ppi |
| [out] | pconf | confidence that text is rightside-up |
| [in] | mincount | min number of up + down; use 0 for default |
| [in] | npixels | number of pixels removed from each side of word box |
| [in] | debug | 1 for debug output; 0 otherwise |
Notes:
(1) See pixOrientDetect() for other details.
(2) The detected confidence conf is the normalized difference
between the number of detected up and down ascenders,
assuming that the text is either rightside-up or upside-down
and not rotated at a 90 degree angle.
(3) The typical mode of operation is npixels == 0.
If npixels > 0, this removes HMT matches at the
beginning and ending of "words." This is useful for
pages that may have mostly digits, because if npixels == 0,
leading "1" and "3" digits can register as having
ascenders or descenders, and "7" digits can match descenders.
Consequently, a page image of only digits may register
as being upside-down.
(4) We want to count the number of instances found using the HMT.
An expensive way to do this would be to count the
number of connected components. A cheap way is to do a rank
reduction cascade that reduces each component to a single
pixel, and results (after two or three 2x reductions)
in one pixel for each of the original components.
After the reduction, you have a much smaller pix over
which to count pixels. We do only 2 reductions, because
this function is designed to work for input pix between
150 and 300 ppi, and an 8x reduction on a 150 ppi image
is going too far -- components will get merged.
Definition at line 558 of file flipdetect.c.
|
static |
Definition at line 209 of file flipdetect.c.
|
static |
Definition at line 215 of file flipdetect.c.
|
static |
Definition at line 221 of file flipdetect.c.
|
static |
Definition at line 227 of file flipdetect.c.
1.8.15