Retrieval of Comic Book Images Using Context Relevance Information

Despite the widespread research interest given in the recent years in analyzing the structure and content of comic books, the question of how to effectively query and retrieve comic images stays a challenge, due to the substantial differences between them and naturalistic images. In this paper, we present a scheme to represent the content in comic-page images using attributed region adjacency graphs. The frequent subgraphs are then mined, and we propose a similarity score for the graphs based on the overlap between them in terms of common component frequent subgraphs. We show that the relationship between the computed similarity score versus panel order can help locating and grouping panels with similar content, or to detect the changing between "scenes", which eventually help to retrieve more relevant results.