Aiming at the problem of blind identification of underdetermined mixtures, an underdetermined blind identification algorithm is proposed, based on the detection of time-frequency single source point and cluster validation technique. Firstly, single source point of each source signal was detected. Then the mixing vector in the corresponding single source point set was estimated by Singular Value Decomposition (SVD). Finally the number of the source signals and the mixing matrix simultaneously were estimated by the cluster validation technique based on k-means clustering algorithm. Compared with the conventional algorithms with single source hypothesis, the proposed algorithm relaxes the sparsity requirement of the source signals and can estimate the mixing matrix under the assumption that there exist disjointed single source points for each source signal. Meanwhile, the proposed algorithm can estimate the number of the source signals while the conventional algorithms require it to be known as a priori. Simulation results display the efficiency of the proposed algorithm. 