| In real life,when consumers purchase clothing products,they often need to go to physical stores to try them on,which consumes a lot of time and energy and also poses public health risks,resulting in poor shopping experiences for consumers.The emergence of virtual try-on technology can solve these problems.It allows consumers to indirectly experience the effect of wearing clothes,judge the compatibility of the selected clothing with themselves,improve the accuracy of consumers’ clothing purchases,and reduce the costs and risks of retailers.However,traditional virtual try-on algorithms generally have problems such as texture loss,large deviation in clothing style,incomplete human posture,and lack of realism.To address these issues,this thesis conducts research on feature-enhanced virtual try-on algorithms based on generative adversarial networks from the perspective of image generation.Firstly,this thesis conducts research on the problem of texture detail loss in virtual try-on algorithm-generated clothing change images.Attention mechanism is introduced into the algorithm network,and the residual block Res Block in the network is replaced with SE-Res Block combined with channel attention to learn the importance of different channel features.Pyramid Squeeze Attention(PSA)module is also introduced in the style encoder to capture spatial information at different scales and enhance the feature representation ability of clothing textures.After a large number of quantitative and qualitative experiments,the proposed improvement scheme in this thesis results in a1.1% increase in SSIM,and the generated image has better image quality and more texture details.Then,based on the improvement of texture details,research is conducted on the problem of generating incomplete human postures and lack of realism.Due to the high-frequency difference between the target posture keypoint image and the background,this thesis introduces Residual Fast Fourier Transform Block(Res FFTBlock)into the posture encoder to capture the global difference between the target posture keypoints and the image background in the frequency domain and enhance the extraction of target posture keypoint features.LPIPS decreased by 1.65%,PSNR increased by 7%.Quantitative experimental results and a large number of qualitative experimental results all proved the effectiveness of the algorithm improvement in this thesis.The generated clothing change images have more complete body structure,more accurate clothing on the human body,and better comply with human visual perception.Finally,based on the trained intelligent try-on model,a virtual dressing web system was designed and developed.The system employs a front-end and back-end separation architecture,utilizing the Vue.js framework and Element UI component library to accomplish front-end development,and utilizing the Flask framework to construct the web service for the back-end.After debugging,integration,and testing,the system achieved the expected effect,realized the virtual try-on function,and demonstrated the practical application effect of this algorithm. |
