Utilizing unique spatial magnetic field information can provide an effective solution for robot localization, especially in repetitive environment. However, it is not trivial to collect magnetic field data at every location due to the single-point sensing limitation of magnetometers and the large scale of environments. Therefore, we introduce the EMFIR method, which efficiently generates dense magnetic field maps from sparsely collected data. This method efficiently processes sparse magnetic field inputs into low-resolution images. Using the continuous image implicit representation, these magnetic field images are then transformed into high-resolution versions, ultimately culminating in the generation of dense magnetic field maps. We evaluated our technique on an indoor self-recorded dataset. Experimental data from real-world environments show that the method proposed in this paper achieves magnetic field mapping with an error of less than 0.05 G, and reduces the RMSE by 76% compared to the traditional Gaussian Process Regression (GPR) method.