// This function takes tif/box pair of files and runs recognition on the image,
// while making sure that the word bounds that tesseract identified roughly
// match to those specified by the input box file. For each word (ngram in a
// single bounding box from the input box file) it outputs the ocred result,
// the correct label, rating and certainty.
void Tesseract::recog_training_segmented(const STRING &fname,
PAGE_RES *page_res,
volatile ETEXT_DESC *monitor,
FILE *output_file) {
STRING box_fname = fname;
const char *lastdot = strrchr(box_fname.string(), '.');
if (lastdot != NULL) box_fname[lastdot - box_fname.string()] = '\0';
box_fname += ".box";
// read_next_box() will close box_file
FILE *box_file = open_file(box_fname.string(), "r");
PAGE_RES_IT page_res_it;
page_res_it.page_res = page_res;
page_res_it.restart_page();
char label[kBoxReadBufSize];
// Process all the words on this page.
TBOX tbox; // tesseract-identified box
TBOX bbox; // box from the box file
bool keep_going;
int line_number = 0;
do {
keep_going = read_t(&page_res_it, &tbox);
keep_going &= read_b(applybox_page, &line_number, box_file, label, &bbox);
// Align bottom left points of the TBOXes.
while (keep_going &&
!NearlyEqual<int>(tbox.bottom(), bbox.bottom(), kMaxBoxEdgeDiff)) {
keep_going = (bbox.bottom() < tbox.bottom()) ?
read_t(&page_res_it, &tbox) :
read_b(applybox_page, &line_number, box_file, label, &bbox);
}
while (keep_going &&
!NearlyEqual<int>(tbox.left(), bbox.left(), kMaxBoxEdgeDiff)) {
keep_going = (bbox.left() > tbox.left()) ? read_t(&page_res_it, &tbox) :
read_b(applybox_page, &line_number, box_file, label, &bbox);
}
// OCR the word if top right points of the TBOXes are similar.
if (keep_going &&
NearlyEqual<int>(tbox.right(), bbox.right(), kMaxBoxEdgeDiff) &&
NearlyEqual<int>(tbox.top(), bbox.top(), kMaxBoxEdgeDiff)) {
ambigs_classify_and_output(page_res_it.prev_word(),
page_res_it.prev_row(),
page_res_it.prev_block(),
label, output_file);
}
} while (keep_going);
}
// This function takes tif/box pair of files and runs recognition on the image,
// while making sure that the word bounds that tesseract identified roughly
// match to those specified by the input box file. For each word (ngram in a
// single bounding box from the input box file) it outputs the ocred result,
// the correct label, rating and certainty.
void Tesseract::recog_training_segmented(const STRING &fname,
PAGE_RES *page_res,
volatile ETEXT_DESC *monitor,
FILE *output_file) {
STRING box_fname = fname;
const char *lastdot = strrchr(box_fname.string(), '.');
if (lastdot != NULL) box_fname[lastdot - box_fname.string()] = '\0';
box_fname += ".box";
// ReadNextBox() will close box_file
FILE *box_file = open_file(box_fname.string(), "r");
PAGE_RES_IT page_res_it;
page_res_it.page_res = page_res;
page_res_it.restart_page();
STRING label;
// Process all the words on this page.
TBOX tbox; // tesseract-identified box
TBOX bbox; // box from the box file
bool keep_going;
int line_number = 0;
int examined_words = 0;
do {
keep_going = read_t(&page_res_it, &tbox);
keep_going &= ReadNextBox(applybox_page, &line_number, box_file, &label,
&bbox);
// Align bottom left points of the TBOXes.
while (keep_going &&
!NearlyEqual<int>(tbox.bottom(), bbox.bottom(), kMaxBoxEdgeDiff)) {
if (bbox.bottom() < tbox.bottom()) {
page_res_it.forward();
keep_going = read_t(&page_res_it, &tbox);
} else {
keep_going = ReadNextBox(applybox_page, &line_number, box_file, &label,
&bbox);
}
}
while (keep_going &&
!NearlyEqual<int>(tbox.left(), bbox.left(), kMaxBoxEdgeDiff)) {
if (bbox.left() > tbox.left()) {
page_res_it.forward();
keep_going = read_t(&page_res_it, &tbox);
} else {
keep_going = ReadNextBox(applybox_page, &line_number, box_file, &label,
&bbox);
}
}
// OCR the word if top right points of the TBOXes are similar.
if (keep_going &&
NearlyEqual<int>(tbox.right(), bbox.right(), kMaxBoxEdgeDiff) &&
NearlyEqual<int>(tbox.top(), bbox.top(), kMaxBoxEdgeDiff)) {
ambigs_classify_and_output(label.string(), &page_res_it, output_file);
examined_words++;
}
page_res_it.forward();
} while (keep_going);
// Set up scripts on all of the words that did not get sent to
// ambigs_classify_and_output. They all should have, but if all the
// werd_res's don't get uch_sets, tesseract will crash when you try
// to iterate over them. :-(
int total_words = 0;
for (page_res_it.restart_page(); page_res_it.block() != NULL;
page_res_it.forward()) {
if (page_res_it.word()) {
if (page_res_it.word()->uch_set == NULL)
page_res_it.word()->SetupFake(unicharset);
total_words++;
}
}
if (examined_words < 0.85 * total_words) {
tprintf("TODO(antonova): clean up recog_training_segmented; "
" It examined only a small fraction of the ambigs image.\n");
}
tprintf("recog_training_segmented: examined %d / %d words.\n",
examined_words, total_words);
}
