Image inpainting based on stacked autoencoders

1st International Scientific School on Methods of Digital Image Processing in Optics and Photonics

物理学;计算机科学

Shcherbakov, O.^1 ; Batishcheva, V.^2

Saint Petersburg National Research University of Information Technologies, Mechanics and Optics, 49 Kronverkskiy pr., Saint-Petersburg

197101, Russia^1

Saint-Petersburg State University, 7-9 Universitetskaya nab., St.Petersburg

199034, Russia^2

关键词: Autoencoders;    Filling in;    Image Inpainting;    Image representations;    Spectrum entropy;    Task-based;   

Others  :  https://iopscience.iop.org/article/10.1088/1742-6596/536/1/012020/pdf
DOI  :  10.1088/1742-6596/536/1/012020

学科分类：计算机科学（综合）

来源: IOP

【 摘 要 】

Recently we have proposed the algorithm for the problem of image inpaiting (filling in occluded or damaged parts of images). This algorithm was based on the criterion spectrum entropy and showed promising results despite of using hand-crafted representation of images. In this paper, we present a method for solving image inpaiting task based on learning some image representation. Some results are shown to illustrate quality of image reconstruction.