New preprint on CNN-based speech balloon detection available

David Dubray and Jochen Laubrock just published a preprint on automatic speech balloon detection and segmentation using deep CNNs, https://arxiv.org/abs/1902.08137. The fully convolutional model, trained on our GNC annotations, achieves state-of-the-art performance on the GNC as well as the eBDtheque data sets. Such semantic segmentation of images is an interesting problem in computer vision and document analysis. Speech balloon (and caption) segmentation can also be considered an important step in building an OCR pipeline for analyzing text in graphic novels.