Some Computational Approaches For Comic Analysis And Understanding

Abstract

Comic research is an important area in academia and industry whose outcomes effectively fuel the multi-billion-dollar entertainment industry. Besides, comics are one of the effective mediums for conveying messages aesthetically since they have ubiquitous influences worldwide in storytelling. Consequently, research on computational approaches to comic analysis is becoming a priority. Motivated by this trend in comic research, this thesis proposes algorithms for developing a robust well-organized system to facilitate mesmerizing experiences for readers on the digital platform. In this thesis, we primarily focus on three open areas of comic research that have received no or less attention to date: (i) de-warping of warped comic document images, (ii) emotion and sentiment analysis of comics, and (iii) comic video summarization. Furthermore, we also proposed some efficient approaches for improving the performance of visual component retrieval from comic document images. However, research on comic analysis still suffers from the deficiency of enough publicly available datasets. Since there is no publicly available comic dataset in Indian languages, we create a new dataset on Bangla comic books that is ab initioquot one among Indian languages. After getting the dataset, we first explore the geometric distortion reduction problem from camera-captured comic document images. Second, we focus on the problem of various visual components (i.e., panels, characters, text boxes, speech balloons, narrative text boxes, and text lines) retrieval from comic images with different shapes and structures. Third, we consider the problem of automatically predicting emotion and sentiment associated with comic scenes, which is still an open area of comic research. Finally, we present an application of the proposed emotion and sentiment analysis methods to the problem of summarization of comic videos.