Balancing content retention in neural style transfer

Abstract

This thesis work focuses on advancements in neural style transfer, a process that enables the blending of content and style features to generate stylized images. It explores feature extraction using two encoders: a VGG19-based encoder and a GLOW based encoder, the latter improving image reconstruction and reducing content leakage through reversible transformations. Various feature fusion techniques are examined, including Adaptive Instance Normalization (AdaIN), Adaptive Attention Normalization (AdaAttN), Self-Attention Network (SANet), Multi-Channel Correlation Network (MCCNet), and Exact Feature Matching, leveraging statistical matching and attention mechanisms. The study also evaluates the impact of different loss functions such as content loss, style loss, identity loss, and contrastive loss on the quality of the output. Custom transformation blocks are introduced, combining methods like feature concatenation, AdaIN with alternative normalizations, and GLOW-based encoders enhanced with attention modules. Existing architectures, such as AdaIN and Exact Feature Matching, are further refined by integrating additional losses to enhance stylization fidelity and preserve content.