Infinitesimal electric and magnetic dipoles are widely used as an equivalent radiating source model. In this paper, an improved method for dipole extraction from magnitude-only electromagnetic-field data based on genetic algorithm and back-and-forth iteration algorithm [1] is proposed. Compared with conventional back-and-forth iteration algorithm, this method offers an automatic flow to extract the equivalent dipoles without prior decision of the type, position, orientation and number of dipoles. It can be easily applied to electromagnetic-field data on arbitrarily shaped surfaces and minimize the number of required dipoles. The extracted dipoles can be close to original radiating structure, thus being physical. Compared with conventional genetic algorithm based method, this method reduces the optimization time and will not easily get trapped into local minima during optimization, thus being more robust. This method is validated by both simulation data and measurement data and its advantages are proved. The potential application of this method in phase retrieval is also discussed.