Scalable image decomposition

Abstract

Image decomposition, which separates a given input image into structure and texture images, has been used for various applications in the fields of computer graphics and image processing. Most previous image decomposition algorithms and applications start with high-quality images, but it requires numerous steps to separate and edit the high-resolution structure and texture images from the low-resolution input counterparts. This paper proposes a simple but effective end-to-end deep neural image decomposition network, which is called ”scalable image decomposition”, by decomposing and upscaling structure and texture images from the degraded input image at the same time. We train the deep neural network to automatically estimate high-resolution structure and texture from the low-resolution input image by designing a shared feature extractor, and structure and texture upscaling networks which have a powerful capability to establish and distinguish a complex mapping between the low-resolution input image and high-quality structure and texture, while preserving more contextual information without any prior information regarding the low-resolution input image. Quantitative and qualitative analyses of the proposed scalable image decomposition network validate that the proposed method is stable and robust against blurring and staircase effects by separating texture and structure upscaling networks in real-time. The predicted upscaled structure and texture images can be used to a variety of applications, such as image abstraction, detail enhancement, and pencil sketching.