Towards open-set text recognition via label-to-prototype learning

• Formulating a new open-set task requires spotting and cognizing novel characters. Proposing framework that handles characters without retraining. fast rectification technique for text recognition. Scene recognition is popular research topic which also extensively utilized in the industry. Although many methods have achieved satisfactory performance close-set challenges, these lose feasibility scenarios, where collecting data or retraining models could yield high cost. For example, annotating samples foreign languages can be expensive, whereas model each time when “novel” character discovered from historical documents costs both resources. In this paper, we introduce formulate demands capability to spot recognize A label-to-prototype learning proposed as baseline task. Specifically, introduces generalizable mapping function build prototypes (class centers) seen unseen classes. An predictor then reject according prototypes. The implementation of rejection over out-of-set allows automatic unknown incoming stream. Extensive experiments show our method achieves promising on variety zero-shot, close-set, datasets.