For every imaging method, optics plays a vital role. Compared to polycapillary optics or a pinhole-collimator, the use of coded apertures as X-ray optics has the advantages of simple fabrication, high sensitivity, and scalability. Therefore, this work explores the feasibility of applying the coded aperture method to X-ray fluorescence imaging. The proposed imaging system consists of a 2D position-sensitive detector coupled to a 2D multi-hole mask, which is parallel and center-aligned to the detector. To reduce the complexity of system calibration and image reconstruction, a separable mask design and a novel near-field coded aperture imaging model were adapted. The performance of the system was investigated using the Geant4 Monte Carlo simulations. Image reconstruction was performed with the iterative algorithm and the deep learning neural network. High quality 2D and 3D images of complex shaped objects can be reconstructed from a single recorded coded image. Unlike imaging systems based on the conventional convolution model, this system can maintain high spatial resolution over a considerable distance range. For the object-to-mask distances of 8 mm and 26 mm, the spatial resolution is 23.7 μm and 36.2 μm, respectively. The 3D reconstruction results show that the system is able to correctly estimate the object-to-mask distance with an axial spatial resolution of 0.75 mm.