How to get tesseract to recognize these alphanumeric strings? - c++

Question

Having a hard time getting tesseract to recognize alphanumeric text. These are not dictionary words, and I've set load_system_dawg and load_freq_dawg to false as recommended in the docs.
Sample code showing what I'm doing:
// Compile with: g++ test.cpp $(pkg-config --libs tesseract opencv)
#include <tesseract/baseapi.h>
#include <tesseract/genericvector.h>
#include <opencv2/opencv.hpp>
int main()
{
/* Disabling the dictionaries Tesseract uses should increase recognition
* if most of your text isn't dictionary words. They can be disabled by
* setting both of the configuration variables load_system_dawg and
* load_freq_dawg to false.
* Source: https://github.com/tesseract-ocr/tesseract/wiki/ImproveQuality
*/
GenericVector<STRING> key;
GenericVector<STRING> val;
key.push_back("load_system_dawg");
val.push_back("false");
key.push_back("load_freq_dawg");
val.push_back("false");
// also tried setting tosp_min_sane_kn_sp to both large and small values
#if 0
// unfortunately, this doesn't work in tesseract v4.0
// https://github.com/tesseract-ocr/tesseract/issues/751
key.push_back("tessedit_char_whitelist");
val.push_back("ABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789");
#endif
std::unique_ptr<tesseract::TessBaseAPI> tess(new tesseract::TessBaseAPI());
tess->Init(nullptr, "eng", tesseract::OEM_DEFAULT, nullptr, 0, &key, &val, false);
tess->SetPageSegMode(tesseract::PageSegMode::PSM_SINGLE_LINE);
for (const auto filename : {"01.png", "02.png", "03.png", "04.png", "05.png", "06.png", "07.png", "08.png", "09.png", "10.png", "11.png", "12.png"})
{
cv::Mat mat = cv::imread(filename);
tess->SetImage(mat.data, mat.cols, mat.rows, mat.channels(), mat.step1());
char *text = tess->GetUTF8Text();
std::string str;
if (text != nullptr)
{
str = text;
}
delete [] text;
std::cout << "OCR results for " << filename << ": " << str << std::endl;
cv::imshow("image", mat);
cv::waitKey();
}
tess->End();
return 0;
}
The images I'm using and the results I'm seeing from tesseract:
"1K 45" (missing whitespace)
"1K"
"ZaP © 13" (top few pixels are missing so this one doesn't surprise me)
"308 8" ('B' vs '8' I understand, but the missing whitespace means we cannot interpret this text)
"23 B 18"
"" (blank!?)
"SZ2EC 5"
"SZ2EC 3"
"1201" (missing whitespace means we cannot interpret these results)
"a)"
"1 € 13"
"J4E 7"
I'm not looking for 100% recognition, but I think there are likely some configuration items in tesseract that I could be setting to get better results than what I'm currently seeing.