Research On Virtual Try-on Technology Based On Deep Learning

In recent years,with the development of deep learning technology,image-based virtual fitting technology has gained more and more attention,and is widely used in online shopping,games,animation and other fields.The virtual fitting method based on thin-plate spline interpolation and mask synthesis is limited by the spatial feature retention performance of the model,which generates irregular clothing deformation in the face of large clothing shape changes,and cannot handle the phenomenon of selfocclusion.In response to these problems,on the basis of researching and analyzing the technology and methods of virtual fitting at home and abroad,this paper has carried out the following work around how to achieve high-quality virtual fitting:(1)In view of the problem that the clothing deformation model is prone to irregular clothing deformation,a perceptual loss is added to the dressing area.In the clothing deformation model,the target clothing is deformed and aligned with the human pose.When calculating the loss of the deformed clothing,first extract the mask of the clothing area in the human body image,and then use the mask to perform a mask operation on the deformed clothing and the human body image to obtain a set of image pairs in the clothing area.,and finally use the VGG19 network to extract the features of the image pair,and perform loss calculation at the feature level.The addition of the loss function can alleviate the problem of irregular deformation of clothing to a certain extent,thereby improving the effect of fitting.(2)In the try-on module,the spatial feature retention performance of the U-Net network is poor,and it is difficult to generate high-quality composition masks,which leads to the problem of self-occlusion,and the U-Net is improved through the cascaded attention mechanism.,the attention mechanism combines channel and spatial attention,channel attention performs weight learning on feature channels,and spatial attention performs weight learning on each pixel position.The attention calculation is performed on the feature map of the skip connection,which enhances the feature retention performance of the network and alleviates the self-occlusion phenomenon of the fitting result.(3)In order to further improve the effect of clothing deformation and improve the self-occlusion processing capability of the model,the fitting module framework has been improved.First,input the human body feature map and the deformed target clothing obtained by the clothing deformation model,and then output the mask of the dressing area,the human image and the further adjusted target clothing through the improved U-Net model,and finally synthesized by the mask.Fitting results.Through the further adjustment of the fitting module,the effect of clothing deformation is improved,and the problem of self-occlusion is effectively solved.This paper conducts experiments on the Zalando dataset.Compared with the results of other virtual fittings,the model designed in this paper can obtain a better generation effect of the clothing area with a small increase in the amount of calculation,and solve the self-occlusion phenomenon well and improve the overall try-on effect.