Blackjack-copilot

Blackjack Copilot recommends a next action(e.g. hit, stand) which has the highest winning rate, that is, optimal for a player.

presentation video

youtube video

To run the code

first do git clone https://github.com/HOchacha/Blackjack-copilot then install the dependency pip install ultralytics opency-python numpy torch PyQt5 Run python3 ./app/util/main.app.py

Description

It gets the card game board image as input, and gives an optimal action as output. Getting a video as input is of course possible since a video is composed of a plenty of continous image frames.

TODO

• Train a YOLOv8 object detection model
• Cluster cards for a dealer and players
• Implement a lookup table of optimal actions
• Recommend action in the case that player has more than two cards
• Make Blackjack Copilot work for video, webcam or stream

Machine Learning Training Result

Object detection is based on YOLOv8s model.

It is trained for 13 epochs(10 epochs are pretrained by PD-Mera).

The training dataset has 10100 original images (except augmented).

Reference

• roboflow Playing Cards Dataset
• GitHub repository as source of pretrained model
• download link of pretrained model
• another GitHub repository which contains the same pretrained model (Jan 27, 2023)
• Ultralytics
• Ultralytics YOLOv8 Docs
• Blackjack playing video (not demo of this project)
• opencv-python