Research on blind face restoration of Tomb Murals based on improved GFPGAN
In order to solve the problems of grave murals such as serious damage,blurred texture and more noise,an E-GFPGAN (Efficiency-Generative Facial Prior Generative Adversarial Network)was proposed,which realizes the blind inpainting of the face part of the tomb mural.Specifically,the StyleGAN3 network is firstly trained to generate the mural faces of the tomb,which provides rich mural face prior information for the network.Then,the skip connection is added to the degradation removal module,and the original convolution is replaced by two layers of dilated convolution to expand the receptive field and retain the detail information of the mural face.Finally,the convolution of the Channel-Split Spatial Feature Transform (CS-SFT)structure was changed into dilated convolution to reduce the number of network parameters.The restoration is carried out on the self-built tomb mural face data set.Experimental results show that the NIQE index is reduced by 2.2％ on average.It is proved that the network can obtain better repair results in the blind restoration of tomb murals.
万方数据
维普
