An improved fixed-point algorithm is used to solve the ICA problem accompanied with noise. According to the noise distribution, it takes two phases to eliminate random noises of different types in preprocessing. The additive white Gaussian noise is removed at first, then the Improved FastICA algorithm is used to process the preprocessed data and to blindly separate the effective signal from non-Gaussian random noise. It might be a problem to set a good starting value in the iterative process of Improved FastICA. In this way, one can accurately set the starting value to make the algorithm recover the effective signal. The satisfactory separation results and better recovery of the effective signal are achieved as shown by the simulation experiments and real seismic data processing. Furthermore, in the case of the strong seismic noise with actually loaded and reduced SNR, this algorithm of blind separation also produces good results. This verifies that the algorithm has good robustness and adaptability. Using the algorithm of blind separation to do the seismic data enhancement can help to better interprete the seismic data, and promote the development of blind separation technology.