OCR

Python OCR using OpenCV

Install

git clone https://github.com/xsyann/ocr.git
cd ocr
tar -xvjf dataset.tbz

Commands:

1. python -m generator.creator [-h] [-p PREFIX] [--dry-run]
2. python -m generator.reader [-h] [-p PREFIX] [--dry-run] filename
3. python generator.py [-h] [-p PREFIX] [--dry-run] [filename]
4. ./test.sh [-h] [-lsd] [-p PATTERN] [filename [filename ...]]
5. python dataset.py [-h] filename
6. python ocr.py [-h] [-lsd] [-t TRAIN_RATIO] [-m MAX_PER_CLASS]

1. Create a dataset entry

python -m generator.creator [-h] [-p PREFIX] [--dry-run]

arguments:

-h, --help      Show help message and exit
-p PREFIX, --prefix PREFIX
                Add a prefix to the filename
--dry-run       Run without writing files

This command open a blank image in which you can draw your letter / symbol / digit. When you have drawn your letter, press the corresponding key (Shift + letter works for majuscules) to write the file dataset/key/prefix-key.i.bmp

• Any Key to write the file
• DEL to remove the last written file
• RET for reset
• UP to increase brush size
• DOWN to decrease brush size
• ESC to quit

Example:

python -m generator.creator -p koeth_y

• Draw an a
• Press a key
• The file dataset/a_small/koeth_y-a_small.0.bmp is written
• Draw another char

---

2. Read a scanned dataset

python -m generator.reader [-h] [-p PREFIX] [--dry-run] filename

arguments:

filename          The file to read
-h, --help        Show help message and exit
-p PREFIX, --prefix PREFIX
                  Add a prefix to the filenames
--dry-run         Run without writing files

This command split a scanned dataset into multiple files.

• SPACE to write the dataset
• UP to increase boxes size
• DOWN to decrease boxes size
• RIGHT to increase image size
• LEFT to decrease image size
• w to move image up
• a to move image left
• s to move image down
• d to move image right

Example:

python -m generator.reader scanned_dataset.bmp -p koeth_y

• Press SPACE

Output:

Write dataset/a_small/koeth_y-a_small.1.bmp
Write dataset/a/koeth_y-a.1.bmp
Write dataset/b_small/koeth_y-b_small.1.bmp
...

---

3. Dataset Generator

python generator.py [-h] [-p PREFIX] [--dry-run] [filename]

arguments:

filename          The file to read
-h, --help        Show help message and exit
-p PREFIX, --prefix PREFIX
                  Add a prefix to the filenames
--dry-run         Run without writing files

If filename is provided execute generator.reader.

Otherwise execute generator.creator.

Examples:

python generator.py -p koeth_y
python generator.py 'scanned_dataset.bmp' -p koeth_y

---

4. Test OCR

./test.sh [-h] [-lsd] [-p PATTERN] [filename [filename ...]]

arguments:

filename      Image to recognize
-h, --help    Show help message and exit
-l            Recognize letters
-s            Recognize symbols
-d            Recognize digits
-p PATTERN, --pattern PATTERN
              Pattern to extract label from filename

If filenames are provided, recognize chars in filenames.

If PATTERN is provided, analyze the predictions.

PATTERN must be a command that extract the classification from the filename.

> pattern.sh 'a_small.jpg'
a

Otherwise, this command open a blank image in which you can draw your letter / symbol / digit. When you press Space, the recognized char appears in a new window.

• SPACE to test
• UP to increase brush size
• DOWN to decrease brush size
• r for rest
• ESC for exit

Examples:

./test.sh -ld `ls dataset/*/*.bmp` -p test/pattern.sh
./test.sh -d
./test.sh -l foo.png
./test.sh -lsd -p test/pattern.sh bar.bmp foo.png

---

5. Display OCR pre-processing

python dataset.py [-h] filename

arguments:

filename      File to pre-process
-h, --help    show help message and exit

Example:

python dataset.py letter.bmp

---

6. Train model

python ocr.py [-h] [-lsd] [-t TRAIN_RATIO] [-m MAX_PER_CLASS]

arguments:

-h, --help      show this help message and exit
-l              Train to recognize letters
-s              Train to recognize symbols
-d              Train to recognize digits
-t TRAIN_RATIO, --train-ratio TRAIN_RATIO
-m MAX_PER_CLASS, --max-per-class MAX_PER_CLASS

TRAIN_RATIO [0.0-1.0] is the ratio between training and test samples (1 = 0 test samples).

MAX_PER_CLASS is the maximum samples (training + test) allowed per class.

Example:

python ocr.py -lsd -t 0.5 -m 50