Neural Style Transfer for Picture with Gradient Gram Matrix Description

Heng Jin , Tian Wang (1) , Mengyi Zhang , Mingmin Li , Yan Wang (2) , Hichem Snoussi (3)

1 Shaanxi University of Science and Technology
2 UC San Francisco - University of California [San Francisco]
3 LM2S - Laboratoire Modélisation et Sûreté des Systèmes

Despite the high performance of neural style transfer on stylized pictures, we found that Gatys et al [1] algorithm cannot perfectly reconstruct texture style. Output stylized picture could emerge unsatisfied unexpected textures such like muddiness in local area and insufficient grain expression. Our method bases on original algorithm, adding the Gradient Gram description on style loss, aiming to strengthen texture expression and eliminate muddiness. To some extent our method lengthens the runtime, however, its output stylized pictures get higher performance on texture details, especially in the elimination of muddiness.

Domains: Engineering Sciences [physics]; Signal and Image processing

Heng Jin, Tian Wang, Mengyi Zhang, Mingmin Li, Yan Wang, et al.. Neural Style Transfer for Picture with Gradient Gram Matrix Description. 2020 39th Chinese Control Conference (CCC), Jul 2020, Shenyang, China. pp.7026-7030, ⟨10.23919/CCC50068.2020.9188652⟩. ⟨hal-03320737⟩