Inpainting Model of Dunhuang Mural Fusing Dynamic Feature and Attention
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.
information processing technologymural inpaintingdeep learningeffective feature selectionattention mechanism
国家科技期刊平台
NSTL
万方数据
