A Virtual Body Try-on Model Based on Edge Detection Algorithm

The target of image visual try-on is to transfer target garment to reference photograph. Current models generally aim to preserve the texture of the target garment while distorting it, and preserve the human features of the non-fitting parts as much as possible. However, when the human image appears occlusion or the human posture changes, it becomes very difficult to fit garment to human body and retain the texture of the garment. In terms of this issue, we present a lightweight network which combines the nonlinear characteristics of convolutional neural network and the characteristics of optimal edge operator. We verified the network on dataset, which achieved good results. Our structural similarity (SSIM) source of warp cloth is 0.8389, SSIM source of human try-on is 0.8130, Frechet Inception Distance (FID) source is 13.6506, all three data are better than mainstream model. Hence, our method can generate fitting images with better perceptual quality and richer detail.