Open-Set Text Recognition Implementations(II): Sample-to-Representation Mapping

This chapter introduces how representations are encoded and extracted from samples, i.e., the sample-to-representation mapping module in the framework discussed above (Fig. 5.1). The contents include three parts: feature extraction, representation aggregation, and context handling. Since most mapping approaches break down into a feature extractor and a sampler module, specifically, the feature extractor maps input images to feature maps, whereas the sampler sample character features from the feature maps. Note for some methods classification modules may be called interleaved, before the sampling procedure, or after the sampling procedure. Despite the execution, the main functionality of the sampler is to aggregate the corresponding regions in the feature map to timestamp aligned targets, so we consider it executes a part of the Sample-to-Representation Mapping functionality. In addition, we discuss the context handling part of the definition of linguistic information and introduce its impact and handling of linguistic information under closed and open environments.