Stepped-frequency ground penetrating radar (GPR) serves as a key non-destructive testing technology for road structure defect detection, where imaging resolution directly affects the reliability and accuracy of defect detection and identification. To address the problems of low imaging resolution in existing stepped-frequency GPR systems for road defect detection, as well as the difficulties in regularization parameter selection and heavy reliance on manual experience when applying compressive sensing methods, a one-dimensional high-resolution imaging method based on the alternating direction method of multipliers (ADMM) network is proposed. The proposed method unrolls the iterative process of the ADMM algorithm into a physically interpretable deep network structure, constructing an end-to-end learning framework consisting of a reconstruction layer, a nonlinear transformation layer, and a multiplier update layer. The reconstruction layer performs backpropagation calculation of signals, the nonlinear transformation layer imposes sparse constraints via a soft-thresholding function, and the multiplier update layer completes the iterative update of Lagrange multipliers. The collaborative work of these three layers enables the network to adaptively learn the optimal parameter combination through training. After obtaining the optimal sequence of reflection coefficients output by the network, it is convolved with a Ricker wavelet to finally generate a high-resolution one-dimensional image. To validate the feasibility of the method, simulation data for three scenarios were collected using gprMax electromagnetic wave propagation simulation software and measured data were collected using the team′s self-developed radar prototype. Simulation and experimental results demonstrate that the proposed method achieves excellent noise immunity while maintaining high resolution. Compared with the improved orthogonal matching pursuit (OMP) algorithm, it improves accuracy by 2% and enhances resolution by approximately two times. When compared to the standard ADMM algorithm, it achieves about a 1% improvement in resolution. These results fully validate the feasibility of the proposed method.