Research On Skeleton Extraction Algorithms Of Calligraphic Characters Based On Deep Learning

The skeleton of calligraphy characters retains the structure,shape and stroke details of calligraphy characters,which is extremely important for evaluating the structure of calligraphy characters.Skeleton extraction technology for calligraphy characters refers to using an algorithm model to refine characters into a single pixel width with complete topological structures.Existing technologies include traditional algorithms and deep learning algorithms,the former is prone to generating distorted skeletons,while the latter is difficult to ensure the refinement and connectivity of the skeleton.To obtain target calligraphic character skeletons,a skeleton extraction algorithm of calligraphy characters with improved conditional generative adversarial network is proposed to learn more complete character structures and richer semantic features.We proposes the Hierarchical Atrous Convolutions Merging module to obtain the long-distance context information.The criss-cross attention module is integrated to ensure the smoothness and accuracy of the generated skeletons.The dataset is a pseudo calligraphy dataset constructed from CASIA-OLHWDB that uses the pen manipulation process as the character skeletons.After training,the network can obtain online information from offline images.In order to further optimize the results of skeleton extraction,an improved network framework of DeeplabV3+is proposed to adapt to the skeleton extraction task,which is mainly achieved by fusing feature information from different scales of the backbone network and using dual attention to integrate deep features to obtain a more complete and smooth character skeleton.The backbone network inheriting the pre-training parameters can effectively accelerate the model training and optimize the index results.Through designing various experiments,the rationality of backbone network selection and the effectiveness of each module are verified.Experimental results show that the proposed network can be generalized to real calligraphy character images and obtain high-quality skeleton extraction results.