In this paper, we describe a hybrid algorithm to create a sketch-like portrait from an input image. First, a user manually selects the face region from the input image. We use an image based method to render the subjects facial features and enhance the shading areas. A stroke based method is then applied to non-face regions to render the silhouette, creases and some dark regions. We show some examples of sketching portraits created by our method and make comparisons to original input images.
Sketching is a classical drawing style that appears in a variety of contexts. Computer generated methods are often used to convert an input image into sketchy styles. But the results are not satisfactory due to the absence of shading variations, vital to express facial information.
Commercial image processing tools concentrate on edge and
silhouette features. Our method includes such features as well, but
focuses on rendering a shading effect inspired by the sketching style
of Leonard da Vinci (as shown in Figure 1) [
We chose human portrait as our sketch subject due to the great interest from the general public. The objective of our sketching style bears some similar characteristics as Figure 1: - Exaggerated shading in the face region; - Curved strokes describing the silhouette and creases; - Straight-line strokes rendering the dark regions; - Background blurred away from the face region with a smooth 2.1
In order to manage the shading and feature information, we use
the tone management approach introduced by Bae et al [
We then recompose the shading enhanced base layer and the feature enhanced detail layer. To simulate the effect of a background blurred away from the face region, we blurred the non-face region as shown in Figure 3. The face region and non-face region of the original image is segmented manually.
The strokes of curves that describe the basic shape such as the silhouette and creases are obtained by Canny edge detection. The process is shown in Figure 4. Straight lines are used to render the shading and to imitate some quick pencil scratches which often appear in many sketch works. More strokes are placed in low intensity regions of the image. The result is shown in Figure 5. 3
Figure 6 shows some results obtained by the above method. For some input images with flat shading such as the first image, we successfully enhance the original flat shading to three distinct shading levels such as the dark shading in the areas of eyes and cheeks, medium dark shading in the nose, and highlights in the forehead and chin. For the input images with more obvious highlight regions and dark regions such as the second image, we keep the original well shaded information (such as the nose region) and enhance some features (such as the eye region).
Our current algorithm still has some drawbacks. The rendering result depends on the quality of the input image. If the input image is of low resolution, the shading and feature information is difficult to enhance. If the input image contains large areas of flat shading, say a large area of pure white or gray, it is hard for the algorithm to enhance the shading into distinct levels. Another drawback is the segmentation of input image into face region and non-face region. In the current implementation, users need to do the segmentation manually; in principle, face recognition could be used to extract a mask automatically.