Progressive Super-resolution Reconstruction Of Dongba Painting Based On Feature Fusion

Dongba painting is an important part of Naxi culture.It is the basis of studying Dongba culture in natural aesthetics,art theory,religion and history.However,the existing low resolution digital images of Dongba painting affect the application,inheritance and development of Dongba culture.Super-resolution reconstruction technology can recover high resolution image from low resolution image.Because Dongba painting has a unique artistic style of rich content,dense lines and diverse colors,when the existing super-resolution algorithm for natural image is directly applied to Dongba painting,the reconstruction effect of Dongba painting’s lines,color blocks and materials is not ideal.Therefore,in order to effectively improve the digital image resolution of Dongba painting,this paper studies the super-resolution reconstruction of Dongba painting,and obtains high-resolution Dongba painting through the reconstruction of low resolution Dongba painting.Firstly,this paper aims at the features of Dongba picture image which contains rich high-frequency information such as edge and texture,and constructs super-resolution reconstruction network combining feature fusion and progressive structure.The overall structure of the network adopts a cascade of multi-level sub networks to reconstruct high-resolution Dongba painting.The reconstruction results of each level of sub network and the labels of corresponding scales are used to calculate the pixel loss.The labels of different scales jointly guide the reconstruction,so as to reduce the loss of high-frequency details in the sampling process of Dongba painting.In each sub network,the shallow feature extraction module and the deep feature extraction module with residual intensive structure are designed respectively to extract the features.Then,the different levels of feature extraction are fused to effectively improve the characteristic loss caused by simple chain stacking of the volume layer.The recursive structure is adopted among all levels of sub networks to share parameters to achieve the balance between parameter quantity and network performance.Secondly,in order to further improve the visual quality of Dongba painting reconstruction,this paper builds a reconstruction network based on the generative adversarial network as the overall framework.The generator adopts the same progressive network structure,which is used to reconstruct low resolution Dongba painting into high resolution Dongba painting;the discriminator is a binary classifier,which is used to judge whether the reconstructed images are true or false.Then,on the basis of pixel loss,perception loss and anti loss are introduced to combat training,which can effectively improve the problems of over smooth reconstruction of Dongba painting and lack of high-frequency details.Finally,in order to make the learning of Dongba painting image features more targeted,this paper constructs a data set DBH2 K containing 298 high-definition Dongba paintings for network training.In DBH2 K,278 of them are used for training and 20 for testing.In this paper,the training set is enhanced by random clipping,and 22024 pieces of 480 × 480 Dongba painting image blocks are obtained for network training.The experimental results show that,compared with Bicubic、SRCNN、LapSRN、Srresnet、IMDN,On the test set of Dongba painting,the proposed algorithm can obtain3.28 db,1.80 db,0.37 db,0.23 db and 0.36 db PSNR gain respectively when the up sampling factor is 8,and the subjective visual quality of reconstructed Dongba painting is also better.The super-resolution network model proposed in this paper can effectively improve the resolution and clarity of low-resolution Dongba paintings,and the network is universal.It can also be trained with other data sets to expand its application range.