Research On Key Technologies Of Calligraphy Image Processing

Rubbings are the most important carrier of calligraphy characters,which is countless and the most authoritative materials for study of calligraphy.However,since many rubbings have been preserved for a long time,they are seriously polluted and damaged,remaining complicated noises,which is hard for image binarization and the characters extraction.In order to solve this problem,this paper proposes rubbing image binarization algorithm based on fully convolutional network,which is summarized as follows:(1)To construct the rubbing binary image resource datasets,this paper has collected ten common rubbing works,segmented out 9430 singlecharacter rubbing images,and manually mark the binary image.(2)This paper designs a fully convolutional network with an encoder-decoder framework.The network inputs the rubbing image and directly outputs the binary image with end-to-end training.(3)The input combines the standard character images and rubbing image,which helps to extract the structural feature of Chinese characters,improving the generalization ability and result of the model.The experimental results show that the network proposed in this paper is very effective and it achieves the state-of-the-art results.Stroke is the basic of characters.Stroke extraction is an indispensable part of the study and evaluation of calligraphy characters.This paper draws on advanced algorithms in the field of instance segmentation and proposes a real-time stroke extraction algorithm for Chinese characters,which divides the strokes into 33 basic strokes.The stroke extraction algorithm consists of two parts.(1)construct a fully convolutional network to generate a set of prototype masks,which will be used to recover stroke instances.(2)add additional prediction head branches to calculate and generate mask coefficients,applying matrix multiplication to linearly combine the prototype masks and the mask coefficient for generating the final stroke instance.In the experiment,the algorithm is tested on the public datasets and the actual rubbings datasets.Compared with other methods,the stroke extraction results of this paper are quite competitive and the efficiency is greatly improved.