Image based human body reconstruction, as a technology for generating realistic human models, can provide fundamental support for virtual reality applications. The current image human body reconstruction algorithms target some subtle human features, such as hand textures, body wrinkles, etc., which are often difficult to accurately capture and reconstruct human details. We will first replace the non-maximum suppression function NMS in the Yolact segmentation model with the Soft-NMS function to better segment the human body in the image. Afterwards, improvements were made to the U-net structure of the Pix2pix backbone network in PIFUHD, which estimates the normal human body graph. The Swin-TransformerV2 module was added to the encoder and decoder structure of U-Net to generate more accurate normal human body graphs. Finally, we validated and tested our method on the Thuman 2.0 dataset, and the experimental results showed that our method improved in evaluation metrics, effectively optimized human reconstruction, and improved the accuracy of human reconstruction.