Oracle Primitive Topology Segmentation Algorithm Based on Mask R-CNN and U-net deep Learning Network

In this paper, considering that the topology is irregular and the data set is small, histogram equalization is used to enhance image contrast, which is conducive to identifying the target boundary, Gaussian blur is used to smooth the image, and image rotation is used to enhance the data. Finally, Sobel operator is used to extract image features from horizontal and vertical directions, and non-local mean method and bilateral filtering are used to discriminate and deal with the identified interference factors. According to the characteristics of the data set and the irregular features of the single word, the data set was normalized and masked by segmentation. The improved Region-based CNN algorithm Mask R-CNN algorithm was used to identify and segment the image, and the single word recognition model was established to make it have good local perception and be able to identify the local features in the image. At the same time, in order to reduce the fitting risk and comprehensively evaluate the model, ten-fold cross-validation is used to improve the training effect of the model. Then, image feature extraction, noise reduction and pixel normalization are carried out on the original topology image, and Ada optimization is carried out on the established single word recognition model to accelerate the convergence of the Mask R-CNN model. Combined with the newly established U-Net model, 200 preprocessed oracle rubbings are trained and segmented.