Digitization efforts by many museums, libraries and other institutions, and the massive growth of both user-generated and professional digital content opens us new possibilities for the study of media and visual cultures. To explore these possibilities, in 2007 we set Software Studies Initiative at University of California, San Diego. Since 2009, our research has been focused on exploring manga culture — specifically, 883 manga titles comprising 1 million digitized pages that were available as scanlations on onemanga.com in the Fall 2009. We have downloaded this whole image set along with the user-assigned tags indicating the genres and intended audiences for these titles and begun its analysis using digital image analysis, visualization, statistics, and data mining techniques.
Our paper presents our methods and some of the key findings of this ongoing research. The discussion is framed by three larger issues: 1) What are the new possibilities for describing visual language of manga made possible by using digital image analysis? 2) What do we learn by combining the software-assisted analysis of individual manga pages and the analysis of user-assigned tags? 3) What does the analysis of large cultural data sets such as 1 million manga pages tells us about cultural categories such as style and genre?
As an example of (1), we present 2D visualizations that show sets of manga pages organized by their visual characteristics measured by software (for instance, the presence of texture/detail and the amount of black in a page). Such visualizations allow us to compare visual language of individual titles, groups of titles, all titles by a particular artist, or 1 million manga pages and large sets of other kinds of images. As another example, we show visualizations that show how visual language of titles changes during the duration of their publication.
To research the relations between visual language, genre and gender/age categories in manga market as manifested in our sample (i.e. issue 2), we analyzed connections between 35 user-assigned tags (4 for audience type — shoujo, shounen, josei and seinen — and 31 for genres) available for 883 titles. We found that all tags form a connected network — i.e. any two tags occur together at least once. While some genre tags are more likely to be assigned to a particular audience segment, none of these genre tags are exclusive to male or female audiences. These empirical findings support two ideas. First, rather than thinking of "action," "adventure," "romance," or any other genre category as a distinct genre, we should instead understand them as genre traits that, according to the perception of manga fans, can be combined in a single title. Second, as constructed by these genre traits, the gender categories female / male strongly overlap.
Digital image analysis and visualization of manga pages supports the similar conclusions. If we organize 1 million manga pages by their visual characteristics and then select all pages corresponding to shoujo and shounen titles, we find that these two sets of pages form overlapping fuzzy "clouds." In other words, while large proportions of pages have distinct visual characteristics that identify them as belonging to shoujo or shounen category, a significant percentage of pages have an "androgynous" visual language.
Finally, we discuss how the analysis of 1 million manga pages leads us to to rethink the concept of style. Consider visualization which shows ourcomplete set of 1 million pages. The pages in the bottom part of the visualization are the most graphic (they have the least amount of detail). The pages in the upper right have lots of detail and texture. The pages with the highest contrast are on the right, while pages with the least contrast are on the left. In between these four extremes, we find every possible stylistic variation. This suggests that manga visual language should be understood as a continuous variable.
This, in turn, suggests that the very concept of style as it is normally used maybe problematic then we consider large cultural data sets. The concept assumes that we can partition a set of works into a small number of discrete categories. However, if we find a very large set of variations with very small differences between them (such as in this case), it is no longer possible to use this model. Instead, it is more appropriate to use visualizations and/or mathematical models to describe the space of possible and realized variations.