Can Deep Learning Approaches Detect Complex Text? Case of Onomatopoeia in Comics Albums

John Benson Louis, Jean-Christophe Burie, Arnaud Revel

Abstract: In recent years, the use of deep learning has contributed to the development of new approaches that outperform traditional methods for many computer vision and document analysis tasks. Many text detection approaches have been proposed and target historical documents, newspapers, administrative documents but also texts in the wild. However, some documents can present complex texts where the shape, the size and the orientation can vary within each word. Onomatopoeia are an example of these complex texts often buried in graphic elements. In this article, we present a study to determine whether the deep learning based text detection methods can tackle such complex texts. First, we selected two well-known deep learning approaches of the literature and then analysed their relevance for the detection of the onomatopoeia. In a second stage, we tried to improve their performance. The experiments show that a simple transfer learning isn’t enough. The models have to consider the characteristics of the onomatopoeia. The proposed strategy show an improvement of about 40% for the detection task of the onomatopoeia.