Paimon Detector

Simple paimon detector using python, opencv (lbpcascade_animeface for detecting faces) and tensorflow.

Project packaged using simple rest api which takes a url to any image and tries to detect whether it has paimon or not.

• Why do you need a paimon detector?
  • Well this is to prevent the recent influx and spam of paimon gif/images on discord.
• Why would someone spam paimon images?
  • I'm not sure ask them.

Setup Instructions

• Run pip install -r requirements.txt
• Run python3 -m uvicorn main:app --reload or python3 main.py

Alternatively you can run the app using docker

• Run docker build -t paimon .
• Run docker run -d -p 8000:8000 paimon

Example Request

curl --location --request GET 'http://localhost:8000/has-paimon?url=https://example.com/some/image.jpg'

Response Format

Will contain an item for each detected face in the image

{
  "faces_detected":true,
  "results":[
    { "possible": "not_paimon", "chances": 99.98887777328491 }
  ]
}

Datasets

• Genshin impact heads
• Anime character face dataset
• Some handpicked images from google

Training results

• Epoch 01: loss: 0.7328 - accuracy: 0.5698 - val_loss: 0.7520 - val_accuracy: 0.4884
• Epoch 02: loss: 0.6573 - accuracy: 0.5756 - val_loss: 0.6219 - val_accuracy: 0.8605
• Epoch 03: loss: 0.5537 - accuracy: 0.8488 - val_loss: 0.4925 - val_accuracy: 0.7674
• Epoch 04: loss: 0.3847 - accuracy: 0.9012 - val_loss: 0.3806 - val_accuracy: 0.8140
• Epoch 05: loss: 0.2630 - accuracy: 0.9070 - val_loss: 0.2487 - val_accuracy: 0.8605
• Epoch 06: loss: 0.2311 - accuracy: 0.9244 - val_loss: 0.1846 - val_accuracy: 0.9302
• Epoch 07: loss: 0.1646 - accuracy: 0.9593 - val_loss: 0.1628 - val_accuracy: 0.9302
• Epoch 08: loss: 0.1503 - accuracy: 0.9477 - val_loss: 0.1206 - val_accuracy: 0.9767
• Epoch 09: loss: 0.1110 - accuracy: 0.9709 - val_loss: 0.1470 - val_accuracy: 0.9302
• Epoch 10: loss: 0.0973 - accuracy: 0.9709 - val_loss: 0.1195 - val_accuracy: 0.9767

Improvements

• Preferably more training data
• Detect more characters?