Multi-scale Gated Fully Convolutional DenseNets for semantic labeling of historical newspaper images

Yann Soullard, Pierrick Tranouez, Clément Chatelain, Stéphane Nicolas, Thierry Paquet

Abstract: Highlights•A new fully convolutional network for historical newspaper semantic labeling.•A multi-scale gate block inspired by dense nets and gating mechanism is proposed.•A multi-scale analysis strengthens the recognition of fields of various sizes.•A gating mechanism provides a focus on features of interest.