The 3D reconstruction process using binocular structured light goes through two streak image processing: wrapped phase solving and phase unfolding. At least 3 images are required to solve the parcel phase using the phase shift method, and acquiring more than one image will affect the efficiency of the reconstruction. In this paper, a neural network model is designed to solve for the wrapped phase using a single high-resolution phase shift image. Firstly, highly accurate wrapped phase results are solved using 12-step phase images and allowed to be used as theoretical reference values for model training. Secondly, using the optimized U-Net end-to-end generation framework, a single phase-shifted image is input, and a two-channel floating-point image of the calculated wrapped phase numerator and the denominator is output. Thirdly, the output two-channel floating-point image is processed with the arctan function to obtain the wrapped phase of the inference output. In the inference phase, the wrapped phase can be solved using a single image using the a priori knowledge learned from the model. The efficiency of structured light 3D reconstruction can be improved by utilizing the high-resolution wrapped phase generation framework proposed in this paper. Compared with the 12-step PS method, the overall reconstruction time is reduced by 87.85%, and the reconstruction accuracy is close to that of the 12-step PS method.