Skip to content

riksantikvarieambetet/Validation-of-object-type-with-machinelearning

 
 

Repository files navigation

Validation-of-object-type-with-machinelearning

Program was created by Daniel Persson and Petter Gullin. This program was created for our B-essay (B-uppsats), the B-essay is included in the repository.

Thank you Albin Larsson (Abbe98 on github) for all the techincal help and guidance on this project

Models and CSV files from our testing

Download link

To use the program

Programfolders Start files
trainingmaterial_app main.py
Trainingtensorflow_scripts create_training_data.py, create_training_data.py
Validation_scripts compare_folders.py, manual_validation.py, validation.py

What is this program?

These programs download, creates and validates picture from Swedish National Heritage Board (Riksantikvarieämbetet) database. This throught API ksamsok.

Why was it created?

These programs were created as the Swedish National Heritage Board saw a problem with the quailty of the metadata.

Purpose of the program

To find wrongfully classifed metadata in their database.

What do the programs do?

Here comes lists about the programs.

General notes

  • The program either validates from a retrained model or a custom trained model.
  • The files in this repository is the two best models from our testing trained on aprox 22000 pictures.
  • The image csv files are before the balancing of the two folders to equal levels.

First program

  • Handels the downloading of the pictures
  • Deleting all corrupted files
  • Removing the info of deleted files from the csv files.

Second program

  • Randomising image files in a folder
  • Creates the trainingdata

Retrained model

  • For the retrained model use this tutorial

Third program

  • Validate against the database by instution with the wrongfully validated pictures getting downloaded
  • Validate manually from either a folder or a file
  • Compare two folders and printout the number of same files in the folders

About

Studentproject created for the Swedish National Heritage Board to help validate metadata of objects and photos in their database throught the help of machinelearning

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%