A Face Sketch Similarity Index for The Evaluation of The Stylized Videos

While the Generative Adversarial Network (GAN) models have been successfully investigated towards the style conversion applications, the existing evaluation metrics suffer from lacking precision in assessing image quality of the style modifications. To alleviate this gap, we have presented a novel Face Sketch Similarity Index (FSSI) technique to enhance the sensitivity of human perception on the quality of the stylized facial videos. In order to identify the optimal parameter combination that greatly enhances model effectiveness, we have systematically explored various training modes and adapting multiple learning rate strategies. The experimental results demonstrated significant improvements on both temporal and spatial performance metrics, which achieved a 10.47 % increasing of Fréchet Inception Distance (FID), a 0.79 % enhancement of FSSI, and a 1.27 % increasing on the inter-frame Structural Similarity Index (SSIM) criterion. Furthermore, by integrating a slight compromise in spatial continuity, we attained more substantial gains, with a maximum improvement by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{2 0. 7 3 \%}$</tex> of FID and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{3. 1 5 \%}$</tex> of FSSI.