Virtual Try-On Network Based On Two-Dimensional Image

The virtual fitting based on two-dimensional image is to transfer the target clothing to the corresponding area of the fitting body,using the strategy of clothing distortion alignment and then fitting synthesis,allowing consumers to experience trying on different clothing online,improving shopping efficiency and consumption frequency,using online fitting shopping as a new way of shopping and consumption growth,making it more convenient and fast for consumers to try on clothes.However,the current method of clothing distortion alignment module for clothing details preservation is not enough and does not take into account the authenticity of the clothing structure,try-on synthesis module for more complex human pose obscuring situation and clothing synthesis of texture details disorder is not solved,therefore,this paper based on two-dimensional image of the virtual try-on need to further research on clothing distortion alignment and try-on synthesis,the main work includes.(1)A multi-stage structure-holding spatial network SPTN that maintains the pixel transformation of the garment design structure combined with the garment structure distortion features is proposed,considering that the geometric matching with thin-slab-like transformation can retain the details but not the design structure of the target garment,and the appearance flow-based method cannot retain the complete details.Therefore,adopt a differential constraint term to smooth the region of deformation excess and use the differential method to constrain the irregular deformation so as to achieve a reasonable distortion of the garment design structure,and propose a strategy of combining pixel transformation and garment structure deformation features first,so that the garment can be more accurately deformed and distorted while maintaining the garment design structure.(2)A semantic prediction module based on human pose and garment feature prediction is proposed.The model aims to separate pose prediction and garment prediction to predict semantic images,and synthesize human pose information and garment details.Considering that previous approaches input rough body shape parameters directly into the network,which leads to confusion between clothing pixels and skin pixels.Therefore,a semantic segmentation prediction model based on body pose and clothing feature prediction is proposed,and a two-stage strategy allows the module to focus more on a single task in order to predict the semantic mapping more accurately after the fitting.(3)A two-encoder based adversarial training synthesis model for convolutional neural network blocks is proposed,which captures and fuses multiple local features with an improved two-encoder class U-Net model and improves the discriminator of the convolutional neural network to implement arbitrary scale cropping of the resulting images by random cropping,thus enabling effective discrimination between the original image and the cropped image.It enhances the data volume,weakens the data noise,has strong generalization ability,improves the model stability,and solves the problem that the existing virtual fitting synthesis methods can have the loss of clothing details and visual artifacts disorder for the case with complex human pose occlusion.Experiments on publicly available datasets show that,compared with conventional analysis methods,The analysis method proposed in this article has a 15% improvement in both qualitative and quantitative research,and its effectiveness is better than the current virtual trial network based on two-dimensional images.