 |
Leptonica 1.85.0
Image processing and image analysis suite
|
|
Leptonica 1.85.0
Image processing and image analysis suite
|
Go to the source code of this file.
Data Structures | |
| struct | L_Recog |
| struct | L_Rch |
| struct | L_Rcha |
| struct | L_Rdid |
Macros | |
| #define | RECOG_VERSION_NUMBER 2 |
Typedefs | |
| typedef struct L_Recog | L_RECOG |
| typedef struct L_Rch | L_RCH |
| typedef struct L_Rcha | L_RCHA |
| typedef struct L_Rdid | L_RDID |
Enumerations | |
| enum | { L_UNKNOWN = 0 , L_ARABIC_NUMERALS = 1 , L_LC_ROMAN_NUMERALS = 2 , L_UC_ROMAN_NUMERALS = 3 , L_LC_ALPHA = 4 , L_UC_ALPHA = 5 } |
| enum | { L_USE_ALL_TEMPLATES = 0 , L_USE_AVERAGE_TEMPLATES = 1 } |
This is a simple utility for training and recognizing individual
machine-printed text characters. It is designed to be adapted
to a particular set of character images; e.g., from a book.
There are two methods of training the recognizer. In the most
simple, a set of bitmaps has been labeled by some means, such
a generic OCR program. This is input either one template at a time
or as a pixa of templates, to a function that creates a recog.
If in a pixa, the text string label must be embedded in the
text field of each pix.
If labeled data is not available, we start with a bootstrap
recognizer (BSR) that has labeled data from a variety of sources.
These images are scaled, typically to a fixed height, and then
fed similarly scaled unlabeled images from the source (e.g., book),
and the BSR attempts to identify them. All images that have
a high enough correlation score with one of the templates in the
BSR are emitted in a pixa, which now holds unscaled and labeled
templates from the source. This is the generator for a book adapted
recognizer (BAR).
The pixa should always be thought of as the primary structure.
It is the generator for the recog, because a recog is built
from a pixa of unscaled images.
New image templates can be added to a recog as long as it is
in training mode. Once training is finished, to add templates
it is necessary to extract the generating pixa, add templates
to that pixa, and make a new recog. Similarly, we do not
join two recog; instead, we simply join their generating pixa,
and make a recog from that.
To remove outliers from a pixa of labeled pix, make a recog,
determine the outliers, and generate a new pixa with the
outliers removed. The outliers are determined by building
special templates for each character set that are scaled averages
of the individual templates. Then a correlation score is found
between each template and the averaged templates. There are
two implementations; outliers are determined as either:
(1) a template having a correlation score with its class average
that is below a threshold, or
(2) a template having a correlation score with its class average
that is smaller than the correlation score with the average
of another class.
Outliers are removed from the generating pixa. Scaled averaging
is only performed for determining outliers and for splitting
characters; it is never used in a trained recognizer for identifying
unlabeled samples.
Two methods using averaged templates are provided for splitting
touching characters:
(1) greedy matching
(2) document image decoding (DID)
The DID method is the default. It is about 5x faster and
possibly more accurate.
Once a BAR has been made, unlabeled sample images are identified
by finding the individual template in the BAR with highest
correlation. The input images and images in the BAR can be
represented in two ways:
(1) as scanned, binarized to 1 bpp
(2) as a width-normalized outline formed by thinning to a
skeleton and then dilating by a fixed amount.
The recog can be serialized to file and read back. The serialized
version holds the templates used for correlation (which may have
been modified by scaling and turning into lines from the unscaled
templates), plus, for arbitrary character sets, the UTF8
representation and the lookup table mapping from the character
representation to index.
Why do we not use averaged templates for recognition?
Letterforms can take on significantly different shapes (eg.,
the letters 'a' and 'g'), and it makes no sense to average these.
The previous version of this utility allowed multiple recognizers
to exist, but this is an unnecessary complication if recognition
is done on all samples instead of on averages.
Definition in file recog.h.
| anonymous enum |
1.12.0