Larger frequency estimation error will appear when a peak spectrum line phase is closed to using frequency interpolation algorithm based on the real part of fast Fourier transform (FFT) coefficient. So an improved sine signal frequency estimation algorithm is proposed. The postion of peak spectrum line is firstly indexed by using the real part and imaginary part sequence of FFT coefficient. And then according to the phase of peak spectrum line, the sequences with larger amplitude in real and imaginary part sequences are selected to participate in frequency interpolation. Simulation results show that at the condition of SNR=3 dB and sampling number N=128, the root mean square of normalized frequency estimation error at the whole frequency band is less than 0.02 and closed to Cramer-Rao lower bound (CRB). The total performance of the improved algorithm is better than the frequency interpolation and Rife algorithm based on the real part of FFT coefficient. The improved algorithm has the characteristics of high frequency estimation precision and few computation amounts and it is easy to implement in hardware.