Letter-Level Online Writer Identification

Writer identification (writer-id), an important field in biometrics, aims to identify a writer by their handwriting. Identification existing writer-id studies requires complete document or text, limiting the scalability and flexibility of realistic applications. To make application more practical (e.g., on mobile devices), we focus novel problem, letter-level online writer-id, which only few trajectories written letters as cues. Unlike text-\(\backslash \) document-based has rich context for identification, there are much fewer clues recognize author from single letters. A main challenge is that person often writes letter different styles time time. We refer this problem variance writing (Var-O-Styles). address Var-O-Styles capture-normalize-aggregate fashion: Firstly, extract features trajectory carefully designed multi-branch encoder, attempt capture styles. Then convert all these style reference feature domain normalization layer. Finally, aggregate normalized hierarchical attention pooling (HAP), fuses input with multiple into compact vector. In addition, also contribute large-scale LEtter-level wRiter IDentification dataset (LERID) evaluation. Extensive comparative experiments demonstrate effectiveness proposed framework.