Detection of Buried Complex Text. Case of Onomatopoeia in Comics Books.

Recent advances in scene text detection, boosted by deep neural networks, have revolutionized our ability to identify text in complex visual environments, including historical documents, newspapers, administrative records, and even text in the wild. However, traditional word-level bounding boxes struggle to capture irregular text shapes. Onomatopoeias in comic books, which epitomize those complex textual elements, are often mixed and buried with graphical components, that make fail conventional detection methods. In this paper, we propose an innovative approach for text segmentation based on the Unet architecture and wisely integrating pre and post-processing. The method is specifically designed to accurately detect intricate text shapes and configurations. This cutting-edge technic offers exceptional accuracy, paving the way for innovative text detection applications in various environments. The evaluations were carried out on the KABOOM-ONOMATOPOEIA dataset and show the relevance of our method in comparison with methods of the literature, which makes it a promising tool in the field of scene text detection.