Abstract: The combination of iterative multiscaling approach (IMA) and deep learning to solve the problem of electromagnetic inverse scattering is of great significance. Based on the IMA framework, this paper proposes a multi-scale iterative unrolling network to solve the nonlinear inverse scattering problem, termed IMA-Net. The network iteratively executes a region of interest (ROI) extraction algorithm, defines an objective function within RoI of different scales, and decomposes the objective function into several optimization sub-problems. The iterative update process of the sub-problems is mapped into the deep network structure, and the relevant components are alternately updated to find the optimal value of the objective function. The results show the accuracy and superiority of proposed method. The studies provide an effective scheme for real-time reconstruction of unknown scatterers.