Open-Set Text Recognition Implementations(I): Label-to-Representation Mapping

This chapter describes the possible approaches to implement the representation-related variable and module in the framework discussed above, i.e., the representation space and the label-to-representation mapping module. First, this chapter introduces how characters, or other corresponding granularities, are represented in different methods, i.e., the representation space, where class centers (prototypes) and features extracted from input images reside. Second, we discuss choices of human representation of labels (side-information) and different approaches in the literature to implement the label-to-representation mapping module. The module, which maps the side information to prototypes residing in the representation space, is the key to implementing class incremental learning functionality.