3D Caricature Reconstruction Based on Shape Features of Single Image
Aiming at the problems of poor landmark detection accuracy and low ability of the generation model to restore high-frequency details in 3D caricature reconstruction from a single image,this paper proposes a two-stage method of multi-scale feature fusion and high-frequency information mapping.In the first stage,a multi-scale channel fusion land-mark detector is used to improve the detection accuracy.The multi-scale features are generated by HRNet,and the atten-tion layer composed of channel attention and Swin Transformer is used for multi-scale channel fusion feature extraction.In order to improve the accuracy of generating landmarks,the loss function consists of two parts:landmark loss and heat map loss.In the second stage,the Fourier feature share layer deformable network enables the generated 3D caricature to have richer high-frequency shape details.Among them,the Fourier feature map extracts high-dimensional features,so that the network can learn more high-frequency information of shapes,and the share layer hypernetwork accelerates the convergence and reconstruction speed of the network.The method is applied to the CaricatureFace and 3DCaricShop datasets.Experimental results show that the average detection error of the landmark detector in this method is reduced by 4.4％;the mean square error of the deformation network in shape reconstruction is reduced by 26％,and the average recon-struction time is reduced by 18％;the final reconstructed 3D caricatures have exaggerated shapes and natural details.
万方数据
维普
