针对目前的算法在细节处理、纹理清晰度以及语义特征连贯性方面存在的问题,提出一种用于人脸图像修复的基于Transformer的U型网络组合部分卷积和空洞卷积模块(Transformer partial convolution and dilated convolution U-shaped network,TPDCU-Net)算法.TPDCU-Net算法将注意力机制中的标准卷积替换为部分卷积,以保留更多可靠信息并降低计算量;同时,在下采样过程中为了减少重要信息的丢失,引入了空洞卷积模块,以改善修复效果.通过在高清人脸(celebfaces attributes high quality,CelebA-HQ)数据集上进行实验,使用峰值信噪比(peak signal-to-noise ratio,PSNR)、结构相似性指数(structural similarity index measure,SSIM)和平均绝对误差(mean absolute error,MAE)指标与现有的图像修复算法进行比较,结果表明当掩码比例最大时TPDCU-Net算法各指标值分别为 23.0493 dB、0.7786、0.0368.该研究证实所提改进算法在人脸图像修复任务中取得了较好的效果.
A TPDCU-Net Algorithm for Facial Image Inpainting
To address the existing issues in current algorithms related to the processing of details,texture clarity,and coherence of semantic features,an algorithm for facial image inpainting based on Transformer partial convolution and dilated convolution U-shaped network(TPDCU-Net)was proposed.In the TPDCU-Net network,standard convolutions in the attention mechanism were replaced with partial convolutions to preserve more reliable information and reduce computational load.Simultaneously,in the downsampling process,dilated convolution modules were introduced to minimize the loss of important information,thereby improving the restoration effectiveness.Experimental evaluations on the celebfaces attributes high quality(CelebA-HQ)dataset compared the proposed algorithm with existing image inpainting algorithms using metrics such as peak signal-to-noise ratio(PSNR),structural similarity index measure(SSIM),and mean absolute error(MAE).The results showed that the corresponding metric values were 23.0493dB,0.7786,and 0.0368,respectively when the mask ratio was maximized.The research indicated that the proposed improvement algorithm achieved better results in facial image inpainting tasks.
万方数据
维普
