To address the issues of insufficient long-range texture dependency modeling, lack of structural coherence, and rough cross-scale feature fusion in mural image restoration, a mural restoration algorithm is proposed based on a two-stage generative adversarial network. Firstly, a multi-scale frequency domain convolution module is designed, using a spatial-frequency parallel architecture and an adaptive expansion rate mechanism to dynamically balance high and low-frequency features and solve the problem of gradual distortion of mineral pigments caused by the limited receptive field of traditional convolution. Secondly, a lightweight bidirectional attention unit is introduced to fuse the direction gradient histogram and sparse attention at the 1/4 resolution layer, enhancing the coherence of large-scale structures. Further, a hybrid attention fusion module is constructed, combining coordinate attention and dynamic channel weighting to optimize multi-scale feature upsampling. Experimental results on the Dunhuang mural dataset show that, compared with the current latest EGMF algorithm, at a mask ratio of [30%, 40%), the peak signal-to-noise ratio (PSNR) index, and the structural similarity (SSIM) index are inproved by 0.13 dB and 1.06%, respectively. And in other 3 kinds of mask ratios, this algorithm outperforms the comparison algorithm, effectively enhancing the visual consistency of the restored image in terms of texture, structure, and details.
CAOJ, ZHANGZ, ZHAOA, et al. Ancient mural restoration based on a modified generative adversarial network[J]. Heritage Science, 2020, 8(1): 1-14.
[3]
吴余芳. 图像修复算法在国画中的应用[D]. 杭州: 浙江大学, 2008.
[4]
SRIDEVIG, SRINIVAS KUMARS. Image inpainting based on fractional-order nonlinear diffusion for image reconstruction[J]. Circuits, Systems, and Signal Processing, 2019, 38(8): 3802-3817.
CAOHaojie, LIYufeng, ZHANGQuan. An improved criminisi fundus image restoration algorithm[J]. Journal of North University of China (Natural Science Edition), 2023, 44(2): 176-181. (in Chinese)
[7]
YAOF. Damaged region filling by improved criminisi image inpainting algorithm for thangka[J]. Cluster Computing, 2019, 22(6): 13683-13691.
[8]
YANGS, LIANGH, WANGY, et al. Image inpainting based on multi-patch match with adaptive size[J]. Applied Sciences, 2020, 10(14): 4921.
[9]
CRIMINISIA, PEREZP, TOYAMAK. Region filling and object removal by exemplar-based image inpainting[J]. IEEE Transactions on Image Processing, 2004, 13(9): 1200-1212.
[10]
KÖHLERR, SCHULERC, SCHÖLKOPFB, et al. Mask-specific inpainting with deep neural networks[M]//Cham: Springer International Publishing, 2014: 523-534.
[11]
GAOL, ZHOUX. A super-resolution reconstruction algorithm for mural images based on improved generative adversarial network[J]. Journal of Measurement Science and Instrumentation, 2024, 15(4): 499-508.
[12]
CHENY, GAOY, ZHUL, et al. PCNet: prior category network for CT universal segmentation model[J]. IEEE Transactions on Medical Imaging, 2024, 43(9): 3319-3330.
[13]
YUJ, LINZ, YANGJ, et al. Generative image inpainting with contextual attention[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018: 5505-5514.
[14]
YUJ, LINZ, YANGJ, et al. Free-form image inpainting with gated convolution[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019: 4470-4479.
[15]
ZENGY, FUJ, CHAOH, et al. Aggregated contextual transformations for high-resolution image inpainting[J]. IEEE Transactions on Visualization and Computer Graphics, 2023, 29(7): 3266-3280.
[16]
CHENW, JIANGZ, WANGZ, et al. Collaborative global-local networks for memory-efficient segmentation of ultra-high resolution images[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019: 8916-8925.
[17]
YUJ, LINZ, YANGJ, et al. Generative image inpainting with contextual attention[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018: 5505-5514.
[18]
DOSOVITSKIYA, BEYERL, KOLESNIKOVA, et al. An image is worth 16x16 words: transformers for image recognition at scale[EB/OL]. [2025-11-03].
[19]
LIW, LINZ, ZHOUK, et al. MAT: mask-aware transformer for large hole image inpainting[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022: 10748-10758.
LIUZhongmin, YANLiang. Inpainting model of Dunhuang mural fusing dynamic feature and attention[J]. Computer Engineering, 2024, 50(5): 342-353. (in Chinese)
[22]
WEIX, FANB, WANGY, et al. Progressive enhancement and restoration for mural images under low-light and defective conditions based on multi-receptive field strategy[J]. NPJ Heritage Science, 2025, 13: 63.
[23]
LEIT, LIY, CHENGZ, et al. Low-rank structure guided diffusion for Shaanxi temple mural restoration[J]. NPJ Heritage Science, 2025, 13: 295.
[24]
LIUZ, LIUY, HUW. Mural restoration via the fusion of edge-guided and multi-scale spatial features[J]. NPJ Heritage Science, 2025, 13: 491.