Research On Recognition Algorithm Of Damaged Inscriptions Based On Deep Learning

The digital preservation of stone inscriptions is of great significance to the historical research and cultural inheritance of inscriptions.However,affected by weathering and man-made destruction,the inscriptions are generally incomplete and ambiguous,and the inscriptions are difficult to recognize and digital preservation is difficult.Due to the lack of prior knowledge and single feature extraction,the current inscription text recognition algorithm results in low recognition accuracy and poor algorithm applicability.Deep learning can automatically extract more advanced and highly distinguished feature information through deepening network for supervised training and learning,and it has high robustness in regular character recognition.Therefore,this paper conducts research on the recognition algorithm of damaged inscriptions based on deep learning.Based on the single-word segmentation of the inscription image by projection segmentation method,the thesis uses artificial synthesis to generate the data set required for training in view of the difficulty in collecting inscription data and the small amount of data.Then a cross-layer convolutional neural network is designed,which improves the utilization rate of features and realizes the prediction and recognition of the single-word inscription after segmentation.Secondly,in view of the problems of inaccurate cutting and long cutting time of the single-word segmentation algorithm,the deep learning network structure of EAST +CRNN is designed.The EAST network is used to achieve pixel-level classification of the processed inscription images,and then the automatically generated target frame is merged and corrected to complete the single-line text segmentation.At the same time,the featureinformation is extracted through CNN in the RCNN network,and the recognition probability of a single line is output using the bidirectional LSTM network,and finally the output results are summarized through the CTC.The recognition method is more convenient and simple,and the recognition speed is faster,and it is effectively verified on the test data set constructed in this paper.