Novelty in Handwriting Recognition

In the domain of Handwriting Recognition (HWR), many novelties may be encountered and can negatively affect the performance of automatic transcription. This is often the case with historical documents. In the context of this unified framework of novelty, HWR is a task consisting of multiple subtasks. The primary task in HWR is transcription, where novel glyphs, characters, words, and phrases may be encountered during the transcription process. Another key task of HWR is style recognition, including writer identification and overall document image appearance. The transcription task involves an agent taking a digital image of a handwritten document as input and processing it to recognize the individual characters to produce a plaintext output. The style recognition task involves the agent identifying known and unknown aspects of visual appearance for both the text, e.g., how are individual characters stylized?, and page i.e., what does the page look like holistically? Two subtasks for style recognition are considered in this paper: (1a) writer identification and (1b) Overall Document Appearance Identification (ODAI). The former involves multi-class classification to distinguish between individual known writers and new writers unseen at training time, while the latter involves multi-class classification to distinguish between known global appearances of handwritten documents and appearances unseen at training time.