Abstract: Based on mean field approximation a new method is proposed to solve noisy ICA model,which can fairly well solve over-complete case,and estimate the independent source by restrict the non-negative mixing matrix and the non-negative sources.The experiments have been done for several different cases,such as digital images,simulated face graphics and ORL face database.The digital images and simulated face graphics experiments show the extraction features by the RMFA-ICA are more independent than that of using tradition ICA and unrestrictive MFA-ICA,the ORL face recognition experiments show the recognition ratio by the proposed method is greater than that of using the others methods.