[🇧🇷] Lê em português

Neural Style Transfer with Neural Networks

This repository contains a Jupyter Notebook that implements the Neural Style Transfer (NST) technique. The goal is to combine the structure of a content image with the artistic appearance of a style image. The method uses the pre-trained VGG19 convolutional neural network to extract and manipulate features from both images.

Project Features

• Neural Style Transfer implementation based on Deep Learning.
• Uses VGG19 for feature extraction.
• Computes the Gram Matrix for style transfer.
• Optimizes the final image using Adam.
• Downloads content and style images during the notebook execution.

Code Structure

1. Loading the VGG19 Network

• Only the features part of the network is loaded.
• The model weights are frozen to prevent retraining.

2. Loading and Processing Images

• The content and style images are loaded.
• Resizing and normalization are applied to ensure compatibility with VGG19.

3. Feature Extraction

• The image is passed through specific convolutional layers of VGG19.
• A mapping of layers to content and style representations is generated.

4. Computing the Gram Matrix

• The Gram Matrix is calculated to measure the correlation between the convolutional filters of the style image.
• This step helps capture the textural patterns of the chosen artwork.

5. Optimizing the Final Image

• The initial image is iteratively adjusted to minimize content and style losses.
• The loss function combines:
  • Content loss (difference between the content representation of the generated and original images).
  • Style loss (difference between the Gram matrices of the generated and style images).
• The Adam Optimizer is used to adjust the pixels of the generated image.

Dependencies

To run this project, install the following libraries:

pip install torch torchvision numpy matplotlib pillow

Execution

To run the Jupyter Notebook, follow these steps:

1. Clone this repository:

   git clone https://github.com/your-username/neural-style-transfer.git
   cd neural-style-transfer

2. Install the dependencies as indicated above.
3. Run the Jupyter Notebook:

   jupyter notebook

4. Open the .ipynb file and execute the cells sequentially.

To run directly on Google Colab, click here

Results

At the end of the execution, the generated image will present the structure of the content image combined with the stylistic features of the style image. This approach is widely used for artistic applications, image filter creation, and AI-generated digital art.

Obtained Result

License

This project is licensed under the MIT License. See the LICENSE file for more details.