Skull Scan

This is a set of software that classifies images using libraries and model from Hugging Face Hub. The first goal is for this software to accurately identify whether an image of a head is fresh or skeletal. The stretch goal is to match a pair of image by identifying if a picture of a skull and a fresh head are of the same person.

To use this you must first make batch folders of the content you want to scan Then you will use the Image_proccessing.py script to make binary forms of the folders Instructions for that can be found here: Usage_of_image_Processing.md

Using the Database

The data we are fine-tuning a model with is located on one of Dr. Mockus's machines and accessible over SSH. You will need to have your UTK VPN activated to access Mockus's machines from home. You will also need to create a ssh key pair. (checkout the git tutorial linked below)

Here are the commands to connect.

• To server the web tool run ssh -p3047 -L3000:localhost:3000 anau@da1.eecs.utk.edu and navigate to localhost:3000 in your web browser.
• To use are docker container run ssh anau@da6.eecs.utk.edu

Docker Details

Our software is set up inside a docker container. hub.docker.com

• To start the image run docker start -a -i skull3 this lets you run commands in the container
• To run the front end server use streamlit run frontEnd.py --server.port=8502 in the container then type http://localhost8502 into your browser
• To create a new container from the image run docker run -it --name skull1 -p8000:8000 mylonjones/huggingface_transformers bash
• To build a new image run docker build -t <image-name> . inside the docker_settings folder
• To exit out of a container run exit
• To get help with docker run docker --help

Extracted data

The shell scripts extract the images from the .gz file and put them into seprate .csv files so we can run the model on this data

• dataExtractor.sh -> takes all stage 4-6 head images and puts into stg4heads.csv
• extractHead3.sh -> takes all stage 1 head images and puts into stg1heads.csv

Resources

Hugging Face Transformers Tutorial

This site has a lot of data sets to work with. kaggle.com

Set up ssh tutorial

Mockus used zcat master_dataset_w_ADD.csv.gz|grep ,head, |cut -d, -f4| sort -R | head -300 > forMylon to copy data from csv file

He used this to remove part of the path on the images sed -i 's|.*/public/||' forMylon

the main csv file is in the default directory master_dataset_w_ADD.csv.gz

forMylon has some head image paths.

start docker docker run -it -p 3050:22 -p 8886:8888 -p 6004:6006 -v /home/anau/:/home/anau/ --name skull huggingface/transformers-pytorch-cpu bash

command to copy files cut -d, -f4 ../stg4heads.csv | grep JPG | while read i; do cp $i .; done

cammand to run example fine-tuner (must be run in image-classification folder) python run_image_classification.py --train_dir ../database --output_dir ../model/ --remove_unused_columns False --do_train --do_eval

Stuff to research use CNNs for image to image top layer SVM a classifier fully connected layer classifies

swap last layer with a classifier

Latest plans of action Tuesday * data script * skull face model research * web ui * github

The Next Tuesday * improve fine-tuining program * testing and metrics * skull-face matching

Final stretch * report * presentation * anything that needs to be done