Oracle Recognition Algorithms Based On Improved PUGAN And CNN Models

The oracle bone inscription can be divided into three categories:handwritten oracle,copied oracle,and rubbing oracle.Rubbing oracle images are rubbing images obtained from turtle and animal bones.Copied oracle images are handwritten by experts based on rubbing oracle images.Handwritten oracle images are high-definition images that has been handwritten by professionals.Handwritten oracle images are the easiest to obtain,while copied oracle images and rubbings are difficult to obtain.Aiming at the problems of few intra-class samples of rubbing oracle,unbalanced samples among classes,and difficulty in obtaining sample identification,an oracle bone inscription recognition algorithm based on improved PUGAN and CNN models was proposed.The oracle bone recognition algorithm based on the improved PUGAN and CNN models achieves the classification from supervised learning of handwritten oracle bones to unsupervised learning classification of rubbing oracle bones through cross-domain adaptation.The main research work of this topic is as follows:(1)Aiming at the problem of low intra-class recognition of similar handwritten oracle,an expansion method of handwritten oracle datasets based on GAN is proposed to increase the generalization ability of the model in the source domain.First,Res Net model is used to recognize the handwritten oracle bones.The sum of the wrong sample data of the statistical model,select similar handwritten oracle bone pairs,and then perform linear interpolation on the latent features.The samples are added to the dataset and retrained,so that the model learns that the image features in the latent space that are not observable by the human eye,improve the classification accuracy of the model on the source domain dataset and the unsupervised recognition accuracy on the target domain dataset.(2)Aiming at the lack of ability of GAN to generate high-quality handwritten oracle images and conduct stable training.On the basis of PUGAN and LSGAN,DPULSGAN model based on the discriminator prior probability distribution of positive examples and unlabeled samples is proposed.The limitations of subjectively selected class priors for PU classification and the feasibility of using the discriminator prior probability distribution as the class prior probability distribution for PU learning.Then,through theoretical analysis,it is proved that minimizing the objective function of the LSGAN network is equivalent to minimizing_dp andp_g Pearson’s chi-square divergence,and DPULSGAN model is equivalent to minimizing the objective function of the DPULSGAN network when ensuring that the generator is in a competitive priority position.The generator dynamic update rule is proposed to keep the generator in the priority position in the confrontation.Finally,through the handwritten oracle dataset generation experiment and the generation experiment on the basic dataset,it is proved that the DPULSGAN model has improved the quality of generated images compared with PULSGAN and LSGAN.The gaussian stability evaluation experiment and similar task stability evaluation experiments show that the DPU algorithm can enhance the training stability of the model.(3)Aiming at the problem that the deep learning method for identifying oracle rubbings lacks a large amount of labeled data and cannot achieve satisfactory performance in recent years,a cross-domain rubbing oracle recognition model based on domain adaptation is proposed.The model first extracts the abstract features of handwritten oracles and rubbing oracles through a feature extractor,and then aligns the features extracted between the source domain and the target domain through an adversarial domain adaptation algorithm.The influence of rubbing images such as severe noise and incomplete images on the accuracy of the model is reduced by minimizing the entropy weight of the target domain.Finally,the multi-pseudo-label algorithm is used to improve the generalization ability of the model and the hit rate of Top-K.