Deep-learning-based image segmentation and text recognition model for oracle original topographies

Oracle bone inscriptions, as the oldest mature form of the Chinese character system, carry the deep heritage of traditional Chinese culture, however, its recognition process is extremely complex and full of challenges. In this paper, we focus on the processing and recognition of the original topographic images of oracle bone inscriptions. First, an image preprocessing model is constructed, which can accurately recognize and effectively remove the noise in the image by virtue of the powerful function of Gaussian filter. At the same time, the model also adopts a non-local homogenization method for different texture features, which further improves the image quality. Then, a Mask R-CNN(Mask Regionbased Convolutional Neural Network) model is introduced to train the image, and a cross-entropy loss function is employed to optimize the model performance, which successfully separates the clear oracle characters from the complex background and interference elements. Finally, an integrated method based on Mask R-CNN segmentation model and VGG(Visual Geometry Group) recognition model is proposed. The method significantly improves the recognition accuracy and efficiency of oracle bone characters through the process of segmentation followed by recognition, which provides a strong technical support for the in-depth development of oracle bone research