| In recent years,the apparel industry,which aims to meet consumers’ shopping needs,has undergone some changes as AR/VR application scenarios have become popular.Realising 3D virtual try-on based on a single image can provide Internet users with more convenient shopping,richer shopping choices and a better shopping experience,which has great commercial value,making it a hot research topic.The detail and realism of generated 3D model textures becomes a major challenge in this field in order to better render results for user satisfaction.The process of virtual 3D try-on based on monocular images can be achieved by fusing 2D information about the human body and clothing and then mapping it onto a 3D body model with depth information.However,most existing methods produce skin-garment penetration after complex garment changes during virtual try-on,excessive distortion during warping clothing,blurred garment prints during try-on;and blurred details such as hand and garment folds and distorted back prediction during depth information prediction.Therefore,this paper proposes a technique for 3D virtual try-on based on monocular images to achieve higher quality 3D virtual try-on based on research and literature analysis.(1)To address the problem of skin-garment penetration and garment and human body misalignment and garment texture blur in the process of 2D information fusion,this paper proposes to reduce the feature loss of location information in low-dimensional space during human body resolution by parallelizing multiple resolutions with constant information interaction between each resolution branch,which effectively alleviates the ambiguous texture generated in the non-garment part due to semantic inaccuracy of human body segmentation;Meanwhile,it proposes to use a constraint module to regularize the TPS parameters to suppress the errors of the TPS parameters caused by overfitting during the warpping clothing;and to use the U-Net with cascaded space and channel attention mechanism for the generation network of the try-on module,where the channel attention and space attention complete the weight learning of the feature channels and each pixel position,respectively,which improves the retention ability of the input garment feature information.The retention ability of the input garment feature information is improved,which in turn improves the detailed representation of the two-dimensional information fusion of the virtual try-on.(2)To address the problem of blurred details in the generated 3D model and ambiguity in the predicted generated occlusion region.By using a parallel multi-resolution approach,more high-frequency features are retained in the pre-processing stage,thus reducing the loss of detail information such as garment folds;at the same time,a multi-scale discriminator architecture is used in the generation network to determine the depth information of the generated occlusion region,and the adversarial loss from the pre-training network is added,thus predicting the generation of more realistic 3D virtual try-on models.With the same dataset,the method in this paper has improved both qualitative comparison and quantitative analysis.Compared with existing related studies,the results generated by this method are rich in detail and realistic,and its superiority is demonstrated by the results of quantitative analysis. |
