Deep-based generation of Wing Interferential Patterns Images for the surveillance of blood-sucking insect population by Machine learning algorithms(Generative adversarial networks, Adversarial Autoencoders).
- Summer intership, research project ETIS LAB, France
This work addresses the task of generating images of wing interference patterns for dataset augmentation within the context of monitoring and accurately identifying diptera specimens to prevent disease spread. These efforts play a crucial role in targeted vaccination campaigns and disease elimination. Current identification methods rely on experts and expensive techniques. The recent work of the ETIS Lab focuses on precise diptera identification using deep learning architectures and Wing Interference Patterns (WIPs), thus mitigating the reliance on these methods. However, challenges involving imbalanced and underrepresented specimens impede the recognition of certain species. This work aims to explore generative architectures that have the potential to enhance the WIPs’ dataset.
Instructions:
# Clone the repository
git clone https://github.com/jhernandezga/Deep-based-generation-of-WIPS.git
# Navigate to the project folder where the document "requirements.txt" is located
# Run: pip install -r requirements.txt
This will install all the necessary packages for the project
# Place the dataset of images in a folder called 'Images': Deep-based-generation-of-WIPS\Resources\Images
```
Use training.py to train a model defining the parameters of the script. For now, training has been mostly tested on images of 256x256 pixels
To visualize training variables at real-time during training using Tensorboard:
- Install Tensorboard for pytorch
pip install tensorboardX - Navigate to the folder of the logger, set when training was initialized
- Execute
tensorboard --logdir path_to_folder --port 6006
Use generate.py to generate images from a trained model
Use tsne.py to run t-SNE algorithm over dataset and generated images
Contact Information: jhernandezga@unal.edu.co