Constraint-Based Reasoning Automatic Generation of Artistic Chinese Calligraphy Songhua Xu, Zhejiang University Francis C.M. Lau, University of Hong Kong William K. Cheung, Hong Kong Baptist University Yunhe Pan, Zhejiang University hinese calligraphy is among the finest and most important of all Chinese art forms C and an inseparable part of Chinese history. Its delicate aesthetic effects are gen- erally considered to be unique among all calligraphic arts. Its subtle power is integral to traditional Chinese painting, where—as figure 1a shows—calligraphy is not just an annotation but also a stylized visual component printing or display; Donald Knuth has done pioneer- affecting the viewer’s emotional response to a paint- ing work in this area.1 We propose an intelligent sys- A novel intelligent ing. This emotional effect also explains why, as fig- tem that can automatically create novel, aesthetically ure 1b shows, calligraphy is often preferred to printed appealing Chinese calligraphy from a few training system uses a type in Asian banners, signage, newspaper mast- examples of existing calligraphic styles. To demon- heads, and other promotional contexts. The 2008 strate the proposed methodology’s feasibility, we have constraint-based Beijing Olympics Games logo (http://en.beijing- implemented a prototype system that automatically 2008.org) is a recent example. generates new Chinese calligraphic art from a small analogous-reasoning Chinese calligraphers predominantly use a soft training set—typically, fewer than 10 samples for each hair brush. Generating artistically appealing callig- character. To the best of our knowledge, no other pub- process to raphy with the brush can be highly challenging. The lished work uses our approach. One remotely related brush-stroke shapes as well as the topology over mul- project uses analogous reasoning to simulate the cre- automatically generate tiple strokes are often very complex. The Chinese ativity in jazz performance and to model other artis- language’s large character set—more than 3,000 tic activities from the simulation.2 original Chinese commonly used characters—presents a problem all its own. Being able to master some of the characters Overall approach and calligraphy that meets doesn’t mean that you can also write the other char- system architecture acters as satisfactorily. Similarly, mastery of one or Let P denote a model with a parameterization E visually aesthetic more styles doesn’t necessarily indicate mastery in that is flexible enough to represent a class of highly other styles, let alone creativity in generating new deformable shapes—different Chinese character requirements. styles. This is where computers can help. styles, in our case. Normally, constructing a flexible In the digital world, calligraphic art is most often model requires significant effort. At the same time, applied to creating typographic or artistic fonts for arbitrary instantiations from such models can easily 32 1541-1672/05/$20.00 © 2005 IEEE IEEE INTELLIGENT SYSTEMS Published by the IEEE Computer Society generate unacceptable results. Thus, a model- based approach to generating novel and yet aesthetically appealing calligraphy is by no means straightforward. Our approach uses a constraint-based analogous-reasoning process (ARP), which we apply to a given set of train- ing examples. Analogous reasoning basically fuses knowledge from multiple sources to sup- port a restricted form of reasoning.3 In our case, the knowledge sources are training examples, which are in the form of images. In our experiments using the prototype system, the training examples come from printed “copybooks” that present multiple calligraphic styles. Because Chinese charac- ters derived from pictographs, which evolved over time into symbols, many basic features recur in different Chinese characters. To take advantage of this redundancy, we devised a hierarchical representation for Chinese char- acters as the basis for our process. The pro- posed ARP consists of three major phases: • Shape decomposition. Decomposing (or recovering) the calligraphic shapes of a (a) (b) given training example is equivalent to the Figure 1. Artistic Chinese calligraphy in Asian societies: (a) Chinese painting with problem of extracting structural features calligraphy; (b) top—ceiling of a Kong Zi (Confucius) temple; bottom—masthead of for constructing a reference model P. The the China Daily newspaper. reference model is an instance of the model P that best represents the input example. The underlying mechanism in models; then, we combine the aligned Figure 2 shows the overall architecture of the our approach is character stroke segmen- models by interpolating or extrapolating proposed intelligent calligraphy generation tation and extraction. the parameterizations {E}. The newly system. • Calligraphy model creation from exam- derived shape family is essentially a “re- ples. Given n reference models {Pi}, parameterization” via the blending para- Character representation where i is the index of the reference model meters, ␸, which control the contribution Our proposed system decomposes Chi- constructed from a set of training exam- of each training example. nese calligraphy into the six levels shown in ples, we can define a family of novel • Artistic calligraphy generation. Given figure 3a: constructive ellipse, primitive shapes P(␸) by blending the reference P(␸) and a set of aesthetics-related geo- stroke, compound stroke, radical, single models: first, we identify the correspond- metrical constraints, we identify some ␸ character, and complete artwork. We adopted ing structural features among the reference that satisfies the given constraints. parametric representations at all levels. Taken Structural stroke database Input images of Shape recovery Hierarchical sample calligraphy through vision representation of from copybook techniques source calligraphy Calligraphy model Calligraphy generation through constraints analogous reasoning Images of Hierarchical Appreciation through Hierarchical automatically representation of Rendering constraint representation of new generated artistic generated artistic satisfaction calligraphy candidates calligraphy calligraphy Figure 2. System architecture for intelligent calligraphy generation. MAY/JUNE 2005 www.computer.org/intelligent 33 CONSTRAINT-BASED REASONING 5th Calligraphy artwork P4, 1 Single- 4th Single characters character level P P P 3rd Radicals 3, 1 3, 2 3, 3 Radical level P P P P P P P P P 2nd Compound strokes 2, 1 2, 2 2, 3 2, 4 2, 5 2, 6 2, 7 2, 8 2, 9 Compound- stroke level P P P P P P P P P P P 1st Primitive strokes 1, 1 1, 2 1, 3 1, 4 1, 5 1, 6 1, 7 1, 8 1, 9 1, 10 1, 11 Primitive- stroke level 0th Constructive ellipses (a) (b) Figure 3. Chinese calligraphy representation: (a) six-level representation hierarchy; (b) four levels in the representation for the Chinese character “zhe.” together, these representations form the para- and the other two rows store the lengths of through quantitative means—analogous meter space E for modeling Chinese callig- its major and minor axes. reasoning together with aesthetic constraint raphy artwork. Traversing the hierarchy from the bottom satisfaction. up, the system first “lines up” the construc- Hierarchical representation enables effi- Prototype implementation tive ellipses to form primitive strokes (level cient local learning of constructive elements In the prototype system, we’ve imple- 1). Then, using shape grammar rules, it com- and reduces the huge global-knowledge mented five typical, frequently occurring bines primitive strokes to form compound representation space to only local shape- primitive strokes: points, horizontals, verti- strokes (level 2), which are subsequently variation characterizations. It also supports cals, left slants, and right slants. Figure 4 combined to form radicals (level 3). By efficient retrieval (and thus reuse) of past shows these strokes as well as 24 typical, fre- grouping radicals according to their spatial calligraphy artwork reasoning results. quently occurring compound strokes and 36 proximity, the system forms single charac- The hierarchical parametric approach radicals. Figure 3b shows the hierarchical rep- ters (level 4). It blends learned examples of can represent all calligraphy styles— resentation of the Chinese character “zhe,” as the same character in different styles into a including cursive styles that are heavily in “Zhejiang,” a scenic coastal province and flexible character model. deformed and distorted—in a uniform six- the home of Zhejiang University. Finally, the top-level constructive element level hierarchy, and it can process the char- Level 0 of the hierarchical representation is calligraphy artwork (level 5), which might acters using the same reasoning pipeline. views an artwork as a collection of con- combine more than one character. This increases our system’s capability to structive ellipses (see figure 3a). The system learn and generate cursive calligraphy, will render the artwork’s “image” as the Advantages of the hierarchical which is an important aspect of calligraphic image space regions that the constructive parametric approach aesthetics. ellipses cover. This representation is inspired Because our approach generates new cal- by the Blum model,4 which defines a zonary ligraphic styles by reasoning from a set of Shape decomposition area by an ellipse moving along a predefined existing styles, it belongs to the hard domain In the first of its three phases, our system curve. A 4 ´ 1