Open-Set Text Recognition: Case-Studies

This chapter surveys two specific open-set text recognition methods, OSOCR and OpenCCD, which support the full feature set of the open-set text recognition tasks, specifically novel character spotting and novel character recognition. The two approaches serve as examples to make case studies of the framework, and may also serve as baseline solutions for the interested readers with the code and models publicly available. For each method, we introduce its motivation, overall design, implementation of individual modules, and training techniques including regularization terms and practical label sampling tricks. This chapter also includes a summarization of the performances of the two models, together with some additional results from the released code.