Sequence Recognition of Scene Text Based on CRNN and CTPN Models

Abstract

Image-based sequence recognition has lately emerged as a prominent study subject in the science of computer vision, while text detection and identification in natural situations has emerged as an active research field. Based on scene text data, this paper addresses the theory of deep learning-based CRNN and CTPN models and the process of processing text. Using CRNN, text recognition can be turned into a time-dependent sequence learning issue, which is commonly employed for indeterminate-length text sequences. Contextual relationships between text images are learned using BLSTM and CTC, thus effectively improving text recognition accuracy and making the model more robust. It also excels in text recognition tests for wordless and lexical-based scenes, as it is not constrained by any predefined language. It produces a more efficient, but smaller, model that is more suited to real-world settings. CRNN recognition accuracy is lower for short texts with large morphological changes, such as artistic words, or texts with large changes in natural scenes. Because of the Anchor setting, CTPN can only detect horizontally distributed text, but a small improvement can detect vertical text by adding horizontal Anchor. As a result of the limitations of the framework, the irregularly inclined text can be detected very broadly.