基于深度学习的楚国墓葬纺织品图像复原

Textile image restoration of Chu tombs based on deep learning

中国楚国时期纺织品由于其墓葬所属年份较为久远,故其纺织品文物在结构和纹样方面存在残缺、破损、污渍等问题。纺织品文物在修复时只能依赖纺织品修复者的经验和审美,且在修复过程中可能会因为修复人员的主观审美及经验出现修复结果不理想或二次损毁等情况。本文通过收集纺织品文物的图像数据建立纺织品图像数据库,采用深度学习方法中的生成对抗网络模型(GAN),针对残缺的纺织品文物,在图像层面进行补全修复,避免了对纺织品文物的接触,减少在修复过程中对文物的二次损毁。通过数字化方法复原纺织品文物的图像,使其结构完整,纹样连贯,从主观评价方面具有较好的复原效果。复原后的纺织品文物图像可以用于指导实物复原、展览展出等,对楚国墓葬纺织品的研究具有一定的借鉴意义。

Most of Chu textile cultural relics in China are unearthed from tombs.Due to the long age and poor storage environment,textiles are often incomplete and damaged with rust spots,blood stains and other blots on the surface.At present,the restoration of textile cultural relics is still mainly physical restoration.The process of physical restoration,depending on the subjective aesthetics and experience of the restoration personnel,is easy to cause problems such as unsatisfactory restoration effect or secondary damage.In order to reduce the contact with textile cultural relics,a digital method was used in this study to restore the textiles of Chu tombs at the image level.The textile image database was established by collecting the image data of textile cultural relics in Chu tombs.The generative adversarial network(GAN)model in the deep learning method was used to repair the incomplete textile cultural relics at the image level,avoiding touching textile cultural relics and reducing the secondary damage in the repair process.At the same time,the generator with the U-Net structure was used in the study to reduce the number of down-sampling and up-sampling,cancel the pooling layer and jump connection,introduce dilated convolution,expand the receptive field,and reduce the loss of textile image data information.According to the characteristics of repeatability and regularity of textile images of Chu tombs,we adopted adaptive step size for the generator.The larger the repair area,the more the details and the more complex the pattern are,and the smaller the step size should be.The smaller the repair area,the less the details and the simpler the pattern are,and the bigger the step size should be,so that the computational efficiency can be improved.The discriminator is of seven-layer network structure,including six convolutional layers and one fully connected layer.The ReLU function was selected as the activation function in the convolutional layers,and the Sigmoid function was selected in the fully connected layer.By learning the data information input into the known area of the textile image database,the generator generated data information close to the unknown area of the real textile image,so as to restore the textile image.The discriminator was used to determine the authenticity of the generated image to adjust the parameters,and the Chu textile image restoration was achieved through continuous optimization of the two.In order to further test the effectiveness of the repair model proposed in this study,the typical image repair method DeepFill v2 algorithm was selected for comparative experiments.The image restored by DeepFill v2 is relatively blurred,with artifacts,noise and other phenomena.The image restored by the method proposed in this study is clear,with the structure and pattern of the textile image being basically restored.In this study,a digital method is proposed to repair some textile cultural relics unearthed from Chu tombs,which are difficult to repair and fragile.The images of textile cultural relics are restored by the digital method to make the structure complete and pattern coherent,which achieves a good restoration effect from subjective evaluation perspective.The restored textile relic images can be used to guide the restoration of physical objects,exhibitions,etc.,which has certain significance for the study of Chu tomb textiles.