Page-level handwritten word spotting via discriminative feature learning

Abstract

Handwritten word spotting (HWS) is a task of retrieving word instances within handwritten documents, which is typically assisted by word annotations (word-level HWS). Previous methods following this paradigm are always in the manual feature modelling fashion, failing to capture sufficient discriminative information of the original input; also, they are always quite time-consuming in the artificial segmentation phase, limiting their applications in practice. To address these problems, we revisit HWS and model it on page-level via discriminative feature learning. Two distinct components modelled as neural networks are combined: word discriminative representation learning by Siamese Feature Network (SFNet) and the word discriminative spotting by Word Discriminative Spotting Network (WdsNet). Even without annotation of boxes, our WdsNet reaches impressive results on the IAM benchmark dataset with 76.8% mAP for the full page word spotting, revealing its superiority over other competitors.