Linear discrimination analysis(LDA) as a classic feature extraction method is widely studied and used. However, LDA as a global criterion is neglected to some extent sample space inner local structure and local information. Therefore, when combined with local weighted mean(LWM) and maximum margin criterion(MMC), a supervised feature extraction method of local learning ability, known as local sub-domains based maximum margin criterion(LBMMC), is proposed. The method is also combined with the QR decomposition technique to improve the efficiency of the algorithm. Finally, the test on the datasets show the effectiveness of the proposed method.