Paolo Andreini (DIISM, University of Siena)
Feb 27, 2019 – 11:00 AM
DIISM, Artificial Intelligence laboratory (room 201), Siena SI
In this seminar, a novel approach to scene text segmentation is presented. The method exploits a new convolutional neural network model, called Segmentation Multiscale Attention Network (SMANet). Employing the SMANet the COCO–Text–Segmentation (COCO TS) dataset, which provides pixel–level supervision for the COCO–Text dataset, is created. The dataset has been automatically generated adopting a weakly supervised approach based on bounding–box annotations. The experiments show that the proposed dataset is a valid alternative to synthetic image generation, which is usually employed for scene text segmentation. The SMANet is trained end–to–end on the proposed dataset, obtaining state–of–the–art results on the ICDAR–2013 segmentation dataset, being surpassed only by multi-stage and task specific architectures.