Open-Set Text Recognition Implementations(III): Open-set Predictor

This chapter discusses the approaches of the representation-prototype matching process, which is used to recognize or reject the corresponding samples in question. For each query instance representation extracted from the sample image, the open-set predictor matches it to representation prototype and determines whether the character belongs to a corresponding class or not. If the instance representation successfully matches a representation prototype, the module yields the label associated with the prototype, otherwise the open-set predictor is rejected, yielding an unknown token. This section studies the inference stage behavior on recognition and rejection, and optimization approaches, specifically loss design and training tricks, that are used to implement the desired behavior.