Research on Scene Chinese Character Recognition Method Based on Similar Chinese Characters

Text recognition in natural scenes has always been a hot topic of research. At present, OCR in academia can support multiple languages and has certain versatility. However, the recognition accuracy of Chinese characters, especially those with similar shapes, is not ideal. Therefore, this paper proposes the Similar-CRNN algorithm based on the traditional CNN + RNN + CTC algorithm model from the perspective of the structure of similar characters and the semantic information of the context. Firstly, we construct a similar character library based on the similarity algorithm of Chinese characters, and conduct enhanced training for the feature differences of similar Chinese characters to improve the recognition accuracy of similar Chinese characters from the aspect of Chinese character structure. Then, after obtaining the preliminary results, add a "semantic detector" to perform three stages of error detection, candidate recall and error correction sorting after Chinese word segmentation, to correct semantically irrelevant error recognition results, and further improve the recognition accuracy rate at the semantic level of Chinese characters.