Recently, advancement in deep learning and three-dimension (3D) imaging technology based on structured light fringe projection on the recovery of the 3D shape of objects from a single fringe image has attracted considerable attention. In this paper, MultiResHNet, an improved global guided path network, is proposed for the 3D shape reconstruction of a single fringe pattern. Herein, the simulation data and experimental data are verified by combining the existing structural optics 3D imaging scheme with the deep convolutional neural network. The experimental results show that the proposed method accurately predicts the 3D shape with lesser errors compared with the existing U-Net neural network. Therefore, our experiments prove the effectiveness and robustness of the proposed method, providing a scientific basis for the improvement of subsequent 3D shape reconstruction with a certain reference value and an application value.