In response to the demand for improving the accuracy of existing backscatter imaging algorithms and the limitations of noise resistance, this paper proposes an inverse scattering imaging method based on the integration of Subspace Optimization Method with Variance Reduction and the M2Net. Within the optimized SOM framework, the Stochastic Variance Reduced Gradient method is introduced, employing a two-layer loop structure that randomly samples a small subset of data in each iteration to update the model, thereby reducing variance through correction terms and improving computational efficiency. Building upon this, a U-shaped nested model called M2Net, incorporating M-shaped residual blocks with multi-scale layers, is constructed. The initial reconstruction results are used as input data for the deep network training of M2Net, enabling further high-precision reconstruction of the scatterer structure. Compared with traditional methods, this method improves structural similarity by 10%~30% and reduces root mean square error by 5%~15%, indicating excellent noise resistance performance and the ability to achieve high-precision image reconstruction.