A new multi-frequency compressed sensing (CS) model is introduced for the 2-D near-field microwave and millimeter-wave synthetic aperture radar (SAR) image reconstruction from under-sampled measurements. The near-filed SAR imaging system usually collects multi-frequency sparse data, where each frequency data can be represented as a hierarchical tree structure under a wavelet basis and different frequency data can be modeled as a joint structure because it's highly correlated. A new multi-frequency CS model, by integrating the tree-sparsity and joint-sparsity together, is proposed to exploit the structural dependencies of the multi-frequency SAR sparse data. In order to solve the corresponding constrained minimization problem, a multi-frequency CS approach is introduced by using a splitting Bregman update with a variation of the parallel Fista-like proximal algorithm. A corrosion-under-paint example demonstrates that the proposed multi-frequency CS model outperforms the conventional model, and the new multi-frequency CS approach enables us to further reduce the number of measurements required to stably recover SAR images of targets and better differentiate true SAR images from recovery artifacts.