Text-Independent Speech Balloon Segmentation for Comics and Manga

Abstract

Abstract-Comics and manga are one of the most popular and familiar forms of graphic content over the world and play a major role in spreading country's culture. Nowadays, massive digitization and digital-born materials allow page-per-page mobile reading but we believe that other usages may be released in the near future. In this paper we focus on speech balloon segmentation which is a key issue for text/graphic association in scanned and digital-born comic book images. Speech balloons are at the interface between text and comic character, they inform the reader about speech tone and the position of its speaker. We present a generic and text-independent segmentation method based on topological and spatial organization of connected component.

Digital Comics Image Indexing Based on Deep Learning

“…Liu et al [32] proposed a clump splitting based speech balloon localization method which can detect both closed and unclosed speech balloons. In addition, Liu et al [33] and Rigaud et al [34] proposed approaches making use of text contained in speech balloons for detecting speech balloon contours at pixel level. At such level of precision, it is also possible to localize and find the direction of the tail.…”

“…In Section 7, we have observed that traditional approaches [31][32][33][34] detect many false positive examples. However, although the deep learning model can localize very well the balloons, it cannot identify precisely the boundaries of the balloons as the traditional approaches.…”

“…Unlike other detection tasks where we only need to detect a rectangle (a bounding box), balloon segmentation requires an algorithm able to detect the boundaries correctly. To boost the performance of balloon extraction, we propose to combine the deep learning approach and a traditional algorithm [34].…”

“…We will use the results of the deep learning segmentation model to remove the FP (False Positive) examples of [34]. This simple technique can also benefit the high boundaries precision of the traditional algorithm and the detected open balloons of the deep learning model to give the best final results (see detail evaluations in Section 7).…”

“…Let D a and D b denote the regions of the balloons detected by [34] and by the deep learning model, respectively. The final regions D f will be computed by the following algorithm 1.…”