Multi-scale sequential network for semantic text segmentation and localization

•New approach for simultaneously localizing and classifying text in presentation slides.•Detected text is classified into three semantic categories: title, bullet, and standard text.•Multi-scale network which extracts features at multiple resolutions and performs semantic text segmentation sequentially.•Bottom-up refinement scheme to perform accurate text localization and semantic text segmentation.•Synthetic dataset of slide images with text and semantic annotation to train the network with abundant data.