There are various outliers which influence the distributions of face samples (signals) and impact the performance of face recognition algorithms. A novel algorithm of kernel independent component analysis for face recognition based on kernel generalized variance and multiresolution analysis (KICA-MKGV) is proposed in this paper. The new algorithm is flexible and robust to a wide variety of signal distributions, and it could extract stable and robust independent features of face samples. According to the experiments on both Harvard face database and FERET face database, the new algorithm could cope with large variation of lighting direction and different illumination intensity very well, and outperform some famous algorithms (PCA, FLD and ICA) in face recognition.