基于局部和全局特征融合的二阶段人脸图像修复算法研究OA北大核心CSTPCD
Research on two-stage face image restoration algorithm based on local and global feature fusion
针对大面积不规则破损的人脸图像修复过程中出现的伪影和不连贯问题,提出一种基于特征融合和多尺度注意力机制的二阶段人脸图像修复算法.在粗修复网络增加全局和局部特征分支来处理编码器的输出.其中,局部特征分支使用多尺度空洞卷积和门控残差连接来聚合上下文信息,并与全局特征分支的输出进行正交融合,提高局部特征与全局特征的相关性,减少特征冗余.在精修复网络增加平均和最大金字塔池化模块,其中,平均池化用于捕捉整体统计信息,最大池化用于提取空间上显著的特征并保留关键信息,并利用通道-空间注意力机制进行图像特征结构调整和纹理生成.最后,构建了一个包括多尺度结构相似性损失的复合函数对网络进行训练.实验结果表明,所提算法在主观和客观评价指标上均优于现有算法.
A two-stage face image restoration algorithm based on feature fusion and multiscale attention mechanism is proposed to address the artifacts and incoherence that occur during the restoration of large irregularly broken face images.Global and local feature branches are added to the rough repair network to process the output of the encoder.Among them,multi-scale dilated convolution and gated residual concatenation are used to aggregate contextual information of the local feature branch,and then the information is orthogonally fused with the output of the global feature branch to improve the correlation between local and global features and reduce the feature redundancy.The average and maximum pyramid pooling modules are added to the fine repair network,among which the average pooling module is used to capture the overall statistical information,and the maximum pooling module is used to extract spatially salient features and retain the key information.In addition,the convolutional block attention module(CBAM)is used for image feature restructuring and texture generation.A composite function including multi-scale structural similarity loss is constructed to train the network.Experimental results show that the proposed algorithm outperforms the existing algorithms in both subjective and objective evaluation indicators.
徐克
山西大学 物理电子工程学院,山西 太原 030006
电子信息工程
全局特征局部特征正交融合金字塔池化CBAM多尺度特征融合人脸图像修复
global featurelocal featureorthogonal fusionpyramid poolingCBAMmulti-scale feature fusionface image inpainting
《现代电子技术》 2024 (009)
40-46 / 7
评论