Uma abordagem híbrida para a renderização não fotorrealística de vídeos digitais.

Abstract

Non-photorealistic rendering (NPR) or stylized rendering is an important growing area of Computer Graphics and Image Processing. NPR is mainly concerned with the generation of images or videos that can have an expressive appeal with regards to the display of features that convey emotions. Some research challenges include: (i) coping with variations in the lighting conditions of the input video; (ii) demanding computational requirements for video rendering; (iii) difficulties in parameter adjustment for each compared approach for a wide range of input videos; (iv) processing videos in a temporally coherent manner; (v) difficulties in the evaluation process. In such a context, this thesis aims at the conception of an original approach to generate stylized videos. The proposed approach was validated by means of a software application, consisting of a six-step pipeline, based on the rendering of each video frame by a modified anisotropic diffusion technique. More specifically, the modifications include: (i) the replacement of the original edge stopping function; (ii) the usage of edges to stylize and guide the diffusion process; (iii) the combination of the detected edges and smoothing by anisotropic diffusion; and (iv) the increase of the saturation and intraregional chromatic homogeneity of the obtained result. A comparative, subjective experimental evaluation of the proposed approach and four concurrent ones was performed, which was based on a user voting scheme. The goal of the evaluation was to highlight improvements in the results of the proposed approach. One-Way ANOVA F-tests followed with multiple means comparison tests, were performed on the results. The proposed approach has obtained better results compared to the four other approaches, regarding the direct visual preference, improved edge definition and stylization. In addition, the results of the proposed approach contain spurious artifacts and flickering which are equivalent with those produced by the compared approaches.