Problem Statement

The project focuses on exploring neural style transfer (NST) to merge artistic styles with content images, addressing the challenge of achieving high-quality style integration while preserving content fidelity. It evaluates various deep learning architectures and loss functions to optimize NST methodologies.

Approach Used

The study establishes VGG19 with Mean Squared Error (MSE) loss as a baseline and compares advanced architectures like ResNet50V2, ResNeXt-50, and DenseNet-121. It also experiments with Adaptive Gram Matrix Loss to dynamically adjust style features for improved aesthetic depth and content balance. Evaluation is based on performance metrics like PSNR and content preservation scores.

Results

• VGG19 with MSE Loss: Baseline with a PSNR of 16.41 dB, providing a foundation for comparisons.
• ResNeXt-50: Achieved the best PSNR of 18.65 dB with excellent content preservation (0.9989).
• Adaptive Gram Matrix Loss: Initial PSNR of 21.15 dB, later declining to 18.07 dB, balancing style and content fidelity.
• DenseNet-121: PSNR of 15.87 dB, offering a good balance between style and content integrity.

Technology Used

• Models: VGG19, ResNet50V2, ResNeXt-50, DenseNet-121.
• Loss Functions: Mean Squared Error (MSE) and Adaptive Gram Matrix Loss.
• Evaluation Metrics: PSNR and content preservation scores for quality assessment.