Image Stylization with Enhanced Structure on GPU

Ping Li,HanQiu Sun,Bin Sheng,JianBing Shen

DOI: https://doi.org/10.1007/s11432-012-4558-y

2012-01-01

Science China Information Sciences

Abstract:This paper presents a graphics processing unit (GPU) based stylization approach that preserves the fine structure between the original and the stylized images using gradient optimization. Existing abstraction and painterly stylization methods focused on contrast manipulation only, and thus the detailed salient structures of the input images are always destroyed when performing the current stylization techniques because of limitations like unavoidable salience information loss caused by contrast abstraction. We propose an image structure map to naturally model the fine structure existing in the original images. Gradient-based structure tangent generation and tangent-guided image morphology are used to construct the structure map. The image structure map, unlike an edge map, not only systematically models the boundary information within the imagery but also accentuates the underlying inner structure detail for further stylization. We facilitate the final stylization via parallel bilateral grid and structure-aware stylizing optimization on a GPU-CUDA platform in real time. In multiple experiments, the proposed method consistently demonstrates efficient and high quality image stylization performance.