Multistage Channel–Spatial Attention Network for Offline Text-Independent Writer Identification

Offline writer identification is used in security fields to identify a writer from a piece of handwriting. It has traditionally concentrated on page-level document images containing multiple paragraphs and sentences, relying on global and local features extracted from entire handwritten documents. In the recent digital age, there has been an increasing demand for word-level writer identification from a minute amount of handwritten text. However, the performance obtained when using the traditional handcrafted features is poor because large amounts of text data are required to obtain sufficient statistical information for writer identification. To alleviate the difficulty and improve performance in offline text-independent writer identification, we proposed a novel multistage channel–spatial attention network. The experimental results using the public CERUG-EN dataset confirm the effectiveness of our proposed method.