Aiming at the problems of low clarity,low face recognition rate and poor visual quality of color avatar images gener-ated from sketch avatars at the present stage,a colorization algorithm for sketch avatars improving CycleGAN was proposed:by optimizing the first feature extraction module of the U-Net self-encoder,a multi-scale self-attention mechanism feature extrac-tion module was designed to extract the input image from multiple scales to reduce the loss of detail information of the input im-age.The extracted features were fused by means of channel stacking,and the fused features were embedded with SENet self-attention mechanism to direct the model's attention to the feature focus area.Finally,the dimension of fused features was reduced.L1 pixel loss and perceptual loss were added to the generated and real avatars to further improve the quality of the generated ava-tars.The experimental results show that compared with the color avatar generated by the base model CycleGAN,the FID value of the CUHK dataset is reduced by 22.23 and Rank-1 value is improved by 16％,and the FID value of the AR dataset is reduced by 15.34 and Rank-1 value is improved by 9.3％.
关键词
CycleGAN/多尺度特征提取/SENet/监督学习/L1像素损失/感知损失
Key words
CycleGAN/multi-scale feature extraction/SENet/supervised learning/L1 pixel loss/perceptual loss
维普
万方数据