Lucs1590 / USeS-BPCA

🧠💻 USeS-BPCA: U-Net Semantic Segmentation BPCA Pooling

USeS-BPCA

U-Net Semantic Segmentation Enhanced with BPCAPooling

Welcome to the USeS-BPCA – a refined repository that brings forward a deep learning model for semantic image segmentation focusing on U-Net architectures. The key spotlight of this repository is the novel BPCAPooling (Block-based Principal Component Analysis Pooling), an advanced dimensional reduction method developed under my Master's thesis that'll soon be up for grabs.

Abstract

The evolution of computer vision has had a significant impact on the efficiencies of image segmentation, particularly in the realms of critical medical analysis, scene analysis, and autonomous system projects. The sharp ascent in the dimensions of the AI world, the arrival of artificial neural networks and deep learning techniques that power multiple architectures are driving state-of-art models offering enhanced performance for different scenarios of image segmentation.

However, the challenge of preserving spatiality when reducing dimensionality, especially in the pooling layers of convolutional networks, continues to be a critical concern. This thesis puts an emphasis on preserving the spatiality of samples during dimensionality reduction, proposing a groundbreaking method known as Block-based Principal Component Analysis Pooling (BPCAPooling). It's a PCA-based pooling method aimed at conserving the spatial structure of the samples and ensuring accurate representations of the learned features for the subsequent neural network layers.

This study embraces the application of BPCAPooling in the convolutional neural network architectures for the classification task, with a primary focus on VGG-16. While the BPCAPooling may not outshine traditional methods in terms of performance metrics for classifications such as accuracy and loss, it manifests as a robust alternative especially in the realm of semantic segmentation yielding a mIoU of $0.3333$, accuracy of $86.77%$ and a loss of $0.6659$.

Architectures

Check out the innovative architectures of U-Nets supplemented with BPCAPooling:

U-Nets with BPCAPooling:

U-Net-Like with BPCAPooling:

Features

• Harness the power of U-Net and U-Net-Like architectures for semantic segmentation;
• Experience BPCAPooling as a viable alternative to conventional pooling methods;
• Explainability of AI with Xplique to understand the model's predictions.

Installation

Ready to explore the USeS-BPCA repository? Follow the steps below:

1. Clone the USeS-BPCA repository:

git clone [e-mail]:Lucs1590/USeS-BPCA.git

2. Install the required packages:

pip install -r requirements.txt

3. Get hands-on with the notebooks folder to run the models.

Contributing

We welcome your contributions. Feel free to raise issues or create pull requests.

License

This project is licensed under the terms of the Apache 2.0 license (see LICENSE for details).

Citation

If you use this repository in your research, please consider citing it. You can cite our work as follows:

@misc{USeS-BPCA,
  author = {Lucas de Brito Silva},
  title = {USeS-BPCA: U-Net Semantic Segmentation Enhanced with BPCAPooling},
  year = {2024},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/Lucs1590/USeS-BPCA}},
}

If you use the BPCAPooling method, please consider citing it. You can cite our work as follows:

@mastersthesis{Silva2024,
  author = {Lucas de Brito Silva},
  title = {Segmentação semântica de imagens com BPCAPooling: uma abordagem baseada em aprendizado profundo},
  year = {2024},
  school = {São Paulo State University (UNESP)},
  address = {Rio Claro, SP, Brazil},
  month = {February}
}

Contact

Got questions? Feel free to reach out to me via email or LinkedIn.