融合结构与纹理特征的壁画缺陷修复

Defect repair of murals guided by fusion structural and textural feature

针对现有算法在修复花纹复杂的壁画时存在结构混乱和纹理模糊等缺陷,提出一种融合结构与纹理特征引导的双重生成对抗网络模型。首先将U-Net引入双重生成网络,利用方向和通道双注意力机制提取到的纹理和结构信息分别引导结构、纹理解码器完成对结构与纹理的特征重构,并结合空洞残差块与跳跃连接实现多尺度特征融合提取。其次将两个分支输出的特征图通过双门控特征融合模块深度融合,完成特征信息交互。最后通过联合双判别器对抗完成缺陷修复,增强壁画修复效果的细节丰富度和全局一致性。实验使用自制数据集五台山某处非国宝级真实壁画进行训练及测试,并通过对比实验和消融实验验证,所提算法在峰值信噪比指标上平均提升4.24 dB,结构相似性指标上平均提升3.6%。实验表明该方法可以对受损的壁画进行有效修复,使其呈现出较好的结构、纹理信息,且视觉效果更清晰自然。

Aiming at the defects of existing algorithms such as structural confusion and texture blurring when repairing murals with complex patterns,a dual-generation adversarial network model incorporating structural and textural feature guidance is proposed.Firstly,U-Net is introduced into the dual-generation network,and the texture and structure information extracted by using the direction and channel dual-attention mechanism guides the structure and texture decoders to complete the feature reconstruction of the structure and texture,respectively,and combines with the null residual block and the jump connection to achieve the extraction of multi-scale feature fusion.Secondly,the feature maps output from the two branches are deeply fused by the dual gated feature fusion module to complete the feature information interaction.Finally,the defect repair is completed through the joint dual-discriminator confrontation,enhancing the detail richness and global consistency of the mural restoration effect.The experiments use self-made dataset of non-national treasure real murals somewhere in Wutai Mountain for training and testing,and verified by comparison experiments and ablation experiments,this paper achieves an average improvement of 4.24 dB in the peak signal-to-noise ratio metric,and improves an average of 3.6%in structural similarity index.The experiments show that the method can effectively repair the damaged murals,so that they present better structural and textural information,and the visual effect is clearer and more natural.