Colored Sketch Portrait Generation System Based On Facial Features Parsing

Portraits are concise yet expressive representations of faces. A life-like portraitshould not only resemble an individual’s appearance, but also the spirit. Due to thefact that human faces vary from one to another, using computer to create a portrait inthe style of particular artistic tradition or an artist is a challenging problem[1].This dissertation introduces an interactive system named Portrait Yourself forgenerating color sketch portrait based on facial feature parsing. It takes a standard IDphoto (front face, without occlusion, in monochrome background) as input andproduces a life-like portrait in a color sketch style.Portrait Yourself contains three main subsystems: Portrait Feature Extraction,Template Matching, and Non-photorealistic Rendering。1. Facial Feature Extraction. We employed an improved ASM[1]for facealignment. This step lays the foundation of portrait generation by providingimportant geometric information. We extract the contours of the hair and clothusing background segmentation and skin color detection.2. Template Matching：a) And-Or graph： In our system we use And-Or graph parsing thevariability of portraits, as well as separating the structure and style. Thisprocedure includes two major steps, each models face and hair-collarrespectively. The face model is further decomposed into7components:eyebrows, eyes, nose, mouth, and face contour. Each component has anumber of distinct sketches (drawn by our artist) as a leaf-node in theAnd-Or graph. The portrait And-Or graph works as a parent templatewhich produces a large set of valid portrait configurations from a set ofsub-templates[3]. b) Template Library. We establish a large template library with detailedclassifications of facial features to generate life-like portrait an artisticstyles. Based on photos we collected from234volunteers, our artist drew177portraits in sketch style. According to the traditional Chineseclassification of facial features, we divided eyebrows into4categories,eyes into6, noses into4, mouths into4, and face contours into6.c) Template Matching. To get the best-match template from the library，weemployed shape context as the main algorithm. We achieve goodmatching result by comparing the shape contexts distance between thetemplate and the primal sketch of the component image.3. Non-Photorealistic Rendering. This system uses key point-based warping togenerate sketch portrait first and then automatically generates light-coloredportrait based on the sketch. Additionally, it also providesbackground replacement and cross stitch style processing as two extrafunctions based on the light-colored portrait.In the end, we summarize our work and discuss about the future work in terms oflimitations and shortcomings of our system.