Improving OCR detection using tesseract

I’m working on a project where I have to detect objects in a PDF document. After detecting the objects, I need to read the text at this location since it will be used as the object's name.

Example:

I’ve managed to detect the objects, I’ve used OpenCV to preprocess the image and want to use Tesseract to read the text from the image.

I’ve used a high resolution image in order to improve Tesseracts accuracy.

I’ve tried using a whitelist, a wordlist and a pattern to further improve tesseracts accuracy. Also, I’ve been playing around with different page segmentation modes, like PSM_SINGLE_WORD and PSM_SINGLE_BLOCK.

Sometimes tesseract reads the text correctly, e.g. the first image returns “T2,T3\n” using PSM_SINGLE_WORD (not using PSM_SINGLE_BLOCK, this returns “12,13\n”). However, in most cases it doesn’t return the correct text.

Preprocessed images for reference:

1st:
Word:  "T2,T3\n"
Block:  "12,13\n"
Expected: “T2,T3\n”

2nd:
Word:  "T2,T3,TAR3\n"
Block:  "12,13, 1T AR3\n"
Expected: “T2,T3,TAR3\n”

3rd:
Word:  "TA8\n"
Block:  "TAR8\n"
Expected: “TAR8\n”

4th:
Word:  "T2TT\n"
Block:  "12,13,14,\nTAR35TAR4\n"
Expected: “T2,T3,T4,\nTAR3,TAR4”

5th:
Word:  "TTT2AA,RRT333A,,\n"
Block:  "12,13,\nTAR35\nTA34\n"
Expected: “T2,T3,\nTAR3,\nTAR34\n”

6th:
Word:  "T15\n"
Block:  "TAR15\n"
Expected: “TAR15\n”

7th:
Word:  "T\n"
Block:  "111\n"
Expected: “T11\n”

As you can see, sometimes PSM_SINGLE_WORD returns better results, sometimes PSM_SINGLE_BLOCK does and sometimes neither returns the correct result.

Since I have quite a few of different variations in the images and I don’t understand why some characters are detected incorrectly (e.g. “,” as “5” in 4th) I’m looking for assistance in resolving this problem.

The relevant code snippet is the following:

Pix* pixImage = pixCreate(eroded.cols, eroded.rows, 8);
for (int y = 0; y < eroded.rows; y++) {
    for (int x = 0; x < eroded.cols; x++) {
        pixSetPixel(pixImage, x, y, eroded.at<uchar>(y, x));
    }
}

QString dataDir = qApp->applicationDirPath() + QStringLiteral("/tessdata");
QString d = QDir::toNativeSeparators(dataDir);

tesseract::TessBaseAPI tess;
QString result;

// First pass: PSM_SINGLE_WORD
if (tess.Init(d.toLatin1(), "eng", tesseract::OEM_DEFAULT) == 0) {
    tess.SetPageSegMode(tesseract::PSM_SINGLE_WORD);
    tess.SetVariable("tessedit_char_whitelist", "TAR0123456789, ");
    tess.SetVariable("user_words_file", "wordList.txt");
    tess.SetVariable("user_patterns_file", "patterns.txt");
    tess.SetVariable("load_system_dawg", "0");
    tess.SetVariable("load_freq_dawg", "0");
    tess.SetVariable("wordrec_enable_assoc", "0");
    tess.SetVariable("use_only_my_words", "1");
    tess.SetImage(pixImage);

    QString wordResult = QString::fromUtf8(tess.GetUTF8Text());
    qDebug() << "Word: " << wordResult;
    result += wordResult;
    tess.End();
}

// Second pass: PSM_SINGLE_BLOCK
if (tess.Init(d.toLatin1(), "eng", tesseract::OEM_DEFAULT) == 0) {
    tess.SetPageSegMode(tesseract::PSM_SINGLE_BLOCK);
    tess.SetVariable("tessedit_char_whitelist", "TAR0123456789, ");
    tess.SetVariable("user_words_file", "wordList.txt");
    tess.SetVariable("user_patterns_file", "patterns.txt");
    tess.SetVariable("load_system_dawg", "0");
    tess.SetVariable("load_freq_dawg", "0");
    tess.SetVariable("wordrec_enable_assoc", "0");
    tess.SetVariable("use_only_my_words", "1");
    tess.SetImage(pixImage);

    QString blockResult = QString::fromUtf8(tess.GetUTF8Text());
    qDebug() << "Block: " << blockResult;
    result += blockResult;
    tess.End();
}

pixDestroy(&pixImage);

Since this is my second question asked on Stack Overflow I might be missing some information, so please feel free to ask for anything you might required to help me.