Darega (experimental)

Machine Learning Neural Network for identifying anime character faces (work-in-progress)

System Requirements

• TensorFlow
• Keras

Training

1. Install TensorFlow and Keras
2. Download training data: animeface-character-dataset prepared by nagadomi. It contains 14490 faces of 177 anime characters.
3. Extract, rename folder "thumb" to "data", put it right next to the python scripts.
4. run python train.py
5. Observe the performance of the model, edit the training script, tune the parameters, change the model, and try again...

Identifying character faces

run python predict.py data\000_hatsune_miku\face_93_104_36.png

It will use the pre-trained model vocaloid.h5, which is trained by using 5 vocaloid character faces: 000_hatsune_miku, 018_kagamine_rin, 033_kagamine_len, 050_megurine_luka, 099_kaito only. It has only trained with 1000 samples in 972 epoch, with limited accuracy.

Sample output

Loading network...
Loading and preprocessing image...
Classifying image...
1/1 [==============================] - 0s
000_hatsune_miku
['86.94%', '0.83%', '0.50%', '0.00%', '4.51%']

Notes

There are many hard-coded parameters in these scripts, make sure to edit and configure them correctly.