Different from CTC-based methods and encoder-decoder-based methods, this chapter proposes a primitive representation learning method that uses global feature aggregation to learn primitive representations from text images. Primitive representations can be regarded as a set of basis vectors in the feature space. Different combinations of primitive representations can generate visual text representations corresponding to the characters to be recognized. Visual text representations can be used for parallel decoding in the implemented primitive representation learning network (PREN). PREN can support both horizontal and vertical text in natural scene images. A semantic-guided decoding method is further incorporated to improve model performance on low-quality images by exploiting both visual and semantic information.