| Chinese calligraphy is the artistic treasure of Chinese traditional culture,and it is a unique visual art,which has high research value.With the digital transformation of museums,it is urgent to use computer technology to protect inscriptions.However,due to the large time span,the poor awareness of human protection in the early stage and the influence of natural weather,the ancient inscriptions have a lot of background noise,which makes it difficult for the traditional digital technology to obtain better visual effect.Therefore,this thesis focuses on the denoising and recognition of ancient inscriptions.Firstly,this thesis introduces the background and significance of the digital processing of ancient inscriptions,and analyzes the research status of image denoising algorithm and recognition algorithm at home and abroad,and focuses on the research work of this paper: establishing the inscription image database;A blind denoising algorithm for tablet image based on multi-scale feature fusion is proposed.An inscription recognition algorithm based on unsupervised classification is proposed.The main researches of the thesis are as follows.(1)For the digitization of ancient inscriptions,there is no large-scale inscription image data base available to the public at present.Therefore,based on the Forest of Steles Museum in Xi ’an,the corresponding data collection standards for different digital tasks of inscriptions are formulated.After collecting inscription image data from multiple angles,and through careful data sorting,screening and classification,a database for inscription image denoising and inscription image recognition is finally established.The database is now available online.(2)In order to solve the problems of inscription image,such as small amount of data and complex and diverse noise modes,a framework of inscription image blind denoising algorithm is proposed.Firstly,an improved GAN model(D-GAN)is used to expand the inscription image data.Based on Bayesian theory,D-GAN learns the distribution of clear images and noisy images,and constructs a large number of noisy images paired with clear images with different noises.Secondly,using the multi-scale feature fusion inscription denoising network model of image denoising,this denoising algorithm by gaussian kernel mode to the original data sampling on multiple scales the image,and use the cycle neural network characteristics of different scale images are extracted,and the characteristics of information fusion for different scales,in order to enhance the details of the denoising image texture information,improve the denoising performance.Finally,aiming at the situation that some strokes were removed in the denoising result of the tablet image,the image restoration algorithm based on coherent semantic attention mechanism was used to restore the denoised tablet image,so as to recover the text information in the denoised tablet image.(3)In view of the different typesetting styles of inscription,the great difference of seal cutting handwriting and the different types of calligraphy in inscription image database,it is difficult to recognize inscription.Therefore,an inscription digital recognition method based on unsupervised classification is proposed.The method consists of two parts: self-coding network and depth-spectrum clustering network.Among them,the feature information of inscription image is learned by using self-coding network;The deep spectrum clustering network embedding these feature information into the feature space for clustering,and the whole clustering algorithm uses KL divergence for collaborative optimization. |
