Fortune On Your Hand: View-Invariant Machine Palmistry

Summary

Our Palmistry principal lines detection software is implemented by 4 steps below. Our main challenge was to read the principal lines on a palm regardless of the view direction and illumination:

1. Warping a tilted palm image
2. Detecting principal lines on a palm
3. Classifying the lines
4. Measuring the length of each line

For palm image rectification, we used MediaPipe to extract interest points and implemented warping with the points. For principal line detection, we built a deep learning model and trained the model with palm image dataset. For line classification, we used K-means clustering to allocate each pixel to specific line. For length measurement, we set a threshold for each principal line with the landmarks obtained by MediaPipe.

Environment

The codes are written based on Python 3.7.6. These are the requirements for running the codes:

• torch
• torchvision
• scikit-image
• opencv-python
• pillow-heif
• mediapipe

In order to install the requirements, run pip install -r ./code/requirements.txt.

Run

1. Before running the codes, a palm image for input(.heic or .jpg) should be prepared in the ./code/inputs directory. We provided four sample inputs.
2. Run read_palm.py by the command below. After running the code, result files will be saved in the ./code/results directory.

> python ./code/read_palm.py --input [filename].[jpg, heic]

Results