| National culture is an important way to spread the spirit of various nationalities.It is of great significance to the production and life of human beings as well as national language.Tie dye is an integral part of national cultural elements.Due to the regional limitation of collection and the difficulty of inheritance,it is facing the problems of data scarcity and inconvenience to its application and research.In addition,the application of some Tie-dye art design to national elements is only formality,which leads to the failure of national style creation to meet the actual needs and lack of use value.This state is not conducive to the protection and development of national Tie-dye elements.This thesis mainly focuses on the technical research of Tie-dye data augmentation and style transfer for the problems of lack of data of national elements and low practicality of creation.The data set selected in the experiment is 2500 pieces from the national culture collection project.After unified processing and selection,129 pieces of standardized tie dye element images can be used as the sample set.The specific work is as follows:(1)An adversarial generation network data augmentation model based on Tie-dye image features as constraint rules is proposed.The model uses a convolutional network to extract the basic texture features of the image and input them into the adversarial generation network.The network structure is optimized so that the generator can generate images with image texture features under the influence of constraint rules.The efficiency of the optimized model is significantly improved.In the experiment,the effect of 1000 to 100000 iterations is preserved.The PSNR value and human eyes discrimination are selected for image quality evaluation.The experimental results show that the image generated by 80000 iterations accords with the characteristics of Tie-dye.Finally,1290 pieces of generated data that reach the image standard are obtained.(2)A style transfer model based on cycle consistency generative adversarial network is proposed.For the problem of image style transfer,the style and content of the image should be separated.In this paper,two discriminators and generators are used to separate the content and style of the image.So the content is retained and does not affect each other.Aiming at the target image where the boundary is blurred,this paper constructs matrix to introduce a mask to help edge optimization,and obtain an image with an improved physical blur style.In the style transfer experiment,1000 scanned images of daily necessities and 1290 data augmentation images were selected as the input sample sets.The migration effect of the experiment is clear with the increase of the number of iterations from 10 to 500,and then tends to be stable.When the number of iterations is from 500 to 1000,the image changes little but the generation efficiency decreases gradually.For the edge improved model experiment,500 iterations are selected to optimize the edge,and the experimental effect is improved obviously.(3)Established a set of Tie-dye image processing and application prototype system.According to the service demand of Tie-dye image data,the image data augmentation and style transfer are performed.This article deploys the system into an application and updates the data processing and feedback in time for subsequent scholars to use.In summary,the method based on the generative adversarial network structure used in this paper can effectively deal with the problem of Tie-dye image data augmentation and style transfer.It has digital protection significance for the new images generated to be transmitted to the cloud for sharing.According to the experimental results in this paper,the image generated by the image data augmentation technology meets the image quality evaluation indicators and the elemental characteristics of Tie-dye.The image generated by the style transfer technology has certain creative significance. |
PDF Full Text Request