A Google image scraper to collect a cleaned and labeled face dataset to train face recognition models. You will need a pretrained model for this project to work. You can either use a text file containing names of use the integrated IMDB name scraper to make the script fully automatic.

In the output folder a text file will be created, containing links to all scraped images and the date they were scraped.

For each person a sub folder name_lastname will be made in the chosen output folder. The faces beloning to the person will be saved in the subfolder as:

name_lastname
	name_lastname_0001.png
	name_lastname_0002.png
	name_lastname_0003.png
	...

Important: if you use your own text file with names make sure the names are formatted as name_lastname.

You can follow the instructions below to deploy this project to your local machine.

For this project to work you first need to install some dependencies. Most dependencies can be installed using pip install -r dependencies.txt. But you will need to install these dependencies manually:

• Facenet
• CUDA
• cuDNN
• Tensorflow-gpu 1.0
• OpenCV 3

Follow the steps below to install and run the project:

1. Clone this repository $ git clone https://github.com/MaartenBloemen/GoogleFaceScraper.git
2. Run the scraper $ python run src/scraper.py /path/to/model.pb /path/to/output/dir/ Optional arguments:
   • --name_source - String - Path to txt file, don't use if you want to use the IMDB scraper
   • --limit - int - Number of IMDB name pages to use (default: 100), number of people are the chosen limit * 50
   • --image_size - int - The width and the height the images will be saved as (default: 160).
   • --margin - int - The margin around the egde of the face and egde of the image (default: 44).
   • --min_cluster_size - int - The minimum amount of pictures required for a single cluster (default: 10), note only the largest cluster will be safed.
   • --cluster_threshold - float - The minimum ecleudian distance for an image to be part of a cluster (default: 1.0).
   • --safe_mode - String - Choices ['on', 'off], this determines whether the Google search should include explicit images or not (default: on).
   • --gpu_memory_fraction - float - A number bewteen 0 and 1 that determines the max percentage of GPU memory that can be used (default: 1.0).

You can find a pretrained model (.pb) on the Facenet repository as well as instructions on how to train your own model.

This project is licensed under the MIT License - see the LICENSE.md file for details.

• Maarten Bloemen

• Facenet - MIT License
• MTCNN - MIT License

This project is meant as a proof of concept as this violates the ToS of Google. To lower the chance of a ban by Google, it's possible to use a VPN.

Do not use datasets gathered this way for commercial use, only for research purposes.

Again this project is purely as a Proof-of-Concept to see what is technical possible.

• Add proxy handler