An application of inpainting-based still image reanimation


Tezin Türü: Yüksek Lisans

Tezin Yürütüldüğü Kurum: Orta Doğu Teknik Üniversitesi, Mühendislik Fakültesi, Elektrik ve Elektronik Mühendisliği Bölümü, Türkiye

Tezin Onay Tarihi: 2017

Öğrenci: ECE SELİN BÖNCÜ

Danışman: GÖZDE AKAR

Özet:

A single-input system that converts still paintings into short video sequences is developed. The system is composed of four separate blocks that are Digital Matting, Inpainting, Motion Modelling & Synthesis and Rendering Blocks. Within the scope of this thesis, the performance analysis of all are realized and the input images are restricted to paintings so as not to cope with the difficulty of achieving photorealism that accompanies the usage of photographs. The input image is partitioned into object-background layers in Digital Matting Block. For the extraction of alpha channels the algorithm proposed in [135] is used and it is tested for performance, especially on highly textured data, which paintings mostly tend to be, coupled with different trimap extraction methods. Inpainting Block is used to cover up the holes in the background after object removal. A detailed analysis and experimentation is done on image inpainting techniques and their performance in the image-to-video converter system is discussed. Within this literature review, a classification based on which key elements are preserved through inpainting is stated, and by experimental verification, it is shown that the geo-texture preserving methods outperformed the methodology based on texture-preservation and geometry-preservation. A further comparison among geo-texture preserving method are made through visual performance and timing analysis in object removal, as well as tests on reconstruction of the known regions of several images, where the abilities of the algorithms in rebuilding the images are tested using several image quality metrics such as PSNR, SNR and SSIM. Motion Modelling & Synthesis Block is responsible for creating an artificial motion field for the object layers. This bi-step procedure involves the sculpting the movement of certain group of objects into mathematical models that are produced by spectral filtering of random noise and generating time dependent displacement maps for each pixel in the image frame. Besides, in order to preserve the harmony within the image frame 2D adaptation of the 3D wind field in [174] is realized. The models implemented within the framework is restricted to passive objects, such as trees, clouds, bodies of water, etc. that move in harmonic oscillations as a reaction to the external drag forces such as wind and gravity. In the ultimate step, Rendering Block, merges the layers back to create the frames of the video by using the transparency maps so as to obtain seamless composite images. The ordering of the frames in the temporal axis provides the flow of the video sequence as the final output.