In this paper, an improved expectation-maximum (IEM) frequency estimation algorithm is discussed, in which the unknown data symbols and the known pilot symbols are exploited jointly in a recursive way. Compared with the data-aided ML estimation, the IEM algorithm can improve the performance without increasing the number of pilot symbols, and is superior in performance to the non-data aided blind estimator at low signal-to-noise ratio (SNR). Furthermore, it avoids converging to a local maximum of the likelihood function rapidly which can deteriorate the estimation performance severely. Simulation results show that IEM estimator, which only use fewer pilot symbols, can provide a good performance close to the CRLB with a large number of pilot symbols at moderately high SNR.