neural_network

Neurons network training (manual)

Prepare training images

• Create a BMP image for each letter of the alphabet (A.BMP, B.BMP, ..., C.BMP).
• Each image must contain a single letter, ideally in black on a white background.
• The images must be of similar size, the code will resize them at 56x49 pixels.

Training via the network for each letter

./train train_single A.bmp
./train train_single B.bmp
./train train_single C.bmp
... et ainsi de suite pour toutes les lettres

Repeat the training step several times for each letter in order to improve precision.

Then, once all the letters are recognized:

./train train dossier_lettres/

Test recognition

./train recognize test_image.bmp

The network automatically saves its condition after each training in "Network_Data.bin" and for the singles in "Train_Data.bin".

---

Network training (automatic) network

Prepare training images

• Choose FONT in TTF format and modify in the generator.py code with the right fact.
• Generate letters in BMP format with letters/generator.py .
• The size of the generated images is defined in generator.py .

Training via the network for each letter

python3 auto_train.py  # exécution une seule fois la première fois

... and wait for the end of training.

Test recognition of each letter

python3 verify_train.py

or for a single letter:

./train recognize test_image.bmp

And if the detected letter is not the right one:

python3 manual_train.py

... then enter the expected letter (ex: U).

Then, to harmonize the network of neurons and save it:

./train train dossier_lettres/

... and so on for all badly detected letters (start the 2 processes again).

The network automatically saves its condition after each training in "Network_Data.bin" and for the singles in "Train_Data.bin".

---

Additional information

• backup_data contains backups of old training sessions.
• python3 manual_train.py allows you to train a single letter.
• python3 auto_train.py is used to train all the letters in a file for the first time.
• python3 verify_train.py is used to check if the letters are well detected.
• train is the version used by the user ( train.c ).
• *.bin contains training data.
• letters/ contains the letters with generator.py .
• checksum.sh is used to save *.bin files.

---

Compilation

gcc train.c -o train -lm ` pkg-config --cflags --libs sdl2 SDL2_image ` 

---

Bookstores to install

• libsdl2-image-dev (for reading images in C).
• PIL (generate letters from Python font).

---

Save the neurons network

bash checksum.sh