Geração de representações não foto-realísticas a partir de fotografias digitais: um estudo de caso com faces humanas.

Abstract

Non-photorealistic rendering (NPR) is an area common to both computer graphics and digital image processing, which focuses on generating simple and comprehensible images from complex scenes, using various levels of abstraction. In this context, sketches are one of the most common and efficient illustration abstractions, being used in many fields, such as animation schemes, entertainment, and advertisement. In this work a technique to automatically enable NPR renderings, which resemble semidetailed sketches of human frontal faces, from digital images is presented and evaluated. The following aspects of the sketch generation have been addressed: (i) pre-processing (color constancy, blurring and histogram transformation) of face images and verification of image quality improvement; (ii) neural network use for edge and boundary detection; (iii) post-processing for improving final results and removing spurious artifacts; (iv) experimental assessment of the produced renderings by objective metrics (PSNR, FoM, SSIM); (v) and validation of the results by an experimental subjective assessment, based on a user voting scheme.