Research on Intelligent Segmentation Method of Oracle Bone Script Images Based on U-Net Model

Abstract

As a precious data for the study of Chinese civilization, Oracle bone inscriptions are difficult to identify due to problems such as rubings, contamination and blur. Although the U-Net deep learning model can effectively retain image details, it still faces challenges such as low contrast, many noise points, complex glyphes and limited data sets when dealing with Oracle bone inscriptions. Gray-scale and binarization processing simplifies the image structure and highlights the text area. At the same time, the mean filter is superior to the Gaussian filter in removing noise, and the PSNR is improved by 15%. The Canny edge detection algorithm achieves an accuracy of 92.3% and effectively removes more than 85% of the interference noise under the optimized parameters. Based on the preprocessing results, a segmentation method based on the U-Net deep learning model is proposed. The model achieves 98.5% pixel classification accuracy on the training set, 89.7% mean intersection over Union (MIU) on the test set, and 91.2% F1 score. It is significantly better than the traditional segmentation method.