Location information of the smart devices (SDs) can be utilized in many applications on the user equipment (UE) side. A moving UE can form a virtual large array (VLA) and receive near-field signals, which can be used to find the SDs' location. Newtonized Orthogonal Matching Pursuit (NOMP), a super-resolution method, estimates the SD's location by applying Newton's refinement (NR) step after OMP estimation. However, OMP, NOMP, and many other methods are generally based on grid search methods and require the appropriate search space. Traditionally, the search space was defined by prior room size information, which is hard to obtain in practice. In this paper, we proposed DNOMP, a NOMP-based method that uses deep learning (DL) to offer the search space. The DL model can significantly reduce the computational complexity but cannot find the accurate location if the training and testing data are from different environments. However, the location error decreases through the NR step. Our three-dimensional ray-tracing simulations show that DNOMP can achieve the same accuracy as other state-of-the-art direct localization methods, but it has less computational complexity and does not require prior information; thus, it can be applied in any indoor environment.