The invention discloses a blind source separation method of decreasing a bit error rate on the basis of ICA. The method specifically includes the following steps: 1) collecting source signal data, andobtaining a mixed signal through linear mixing; 2) whitening the mixed signal, which is obtained in the step 1), to obtain a whitened matrix; 3) calculating bit error rates of source signals in the step 1); 4) selecting an initial weight vector, and using maximum likelihood estimation as a criterion to construct an iteration function; 5) combining the bit error rates, which are obtained in the step 3), to normalize the initial weight vector obtained in the step 4); 6) judging whether the normalized initial weight vector converges, if yes, executing a step 7), and otherwise, continuing iteration until the same converges; and 7) verifying a separation effect on signals after unmixing from angles of correlation coefficients and the system error rate after algorithm execution ends. The methodenables image signal separation to be high in precision and good in convergence.