融合动态特征与注意力的敦煌壁画修复模型

Inpainting Model of Dunhuang Mural Fusing Dynamic Feature and Attention

敦煌壁画图像含有丰富的纹理和结构信息,在破损壁画修复过程中,容易忽略受损特征信息与完整特征信息之间的区别,从而误导修复过程生成不合理的壁画内容。针对该问题,提出一种融合动态特征选择和像素级通道注意力的壁画修复模型。设计基于U-Net的网络生成器,实现对破损图像的编码与解码操作;采用有效可迁移卷积模块,通过动态选择采样空间位置,实现对有效特征信息的灵活提取,采用区域综合归一化模块减少修复区域与完整区域的期望和方差的偏移,从而加强对有效特征信息的选择和利用;在解码层设计像素级通道注意力模块,在增强有效特征权重的同时使模型可从相隔较远的空间位置学习有效特征。在敦煌壁画数据集上的实验结果表明,该算法能够利用有效信息修复掩膜区域比例不一的不规则破损壁画图像,相比PConv、PRVS、DSNet算法,在峰值信噪比(PSNR)指标上平均提升0.502 d B,在结构相似性(SSIM)指标上平均提升1.384%。

Dunhuang mural images contain abundant textural and structural information.During the repair of inpainting murals,differences between damaged and complete feature information are easily overlooked,which can result in unsatisfactory repair and unreasonable mural content.Hence,a mural-restoration model that integrates dynamic feature selection and Pixel-Level Channel Attention(PLCA)is proposed in this study.It designs a U-Net-based network generator to encode and decode inpainting images as well as adopts an effective transferable convolution module that dynamically selects the sampling-space position to flexibly extract effective feature information.A region synthesis-normalization module is used to reduce the expected and variance deviations between the repaired and complete areas,thereby strengthening the selection and utilization of effective feature information.Finally,a PLCA module is designed in the decoding layer to enhance the effective feature weights while allowing the model to learn effective features from distant spatial positions.Experimental results on the Dunhuang mural dataset show that the proposed algorithm can effectively use information to repair irregularly damaged mural images with varying proportions of mask areas.Compared to PConv,PRVS,DSNet algorithms,the proposed algorithm improves the Peak Signal-to-Noise Ratio(PSNR)and Structural Similarity(SSIM)index by 0.502 dB and 1.384%on average,respectively.