According to the analysis on characteristics of blind pixels,a method of taking the fixed pattern noise and random spatial-temporal noise in 3D noise model as data source for blind-pixel detection is proposed,and the feature vector of pixels is formed.The bivariate normality distribution model of feature vector is constructed and the blind pixels are regarded as abnormal pixels,dissociating from the normal distribution ellipsoid.The statistical distance in the feature space and the feature space angle are regarded as statistical criteria of detection of abnormal pixels,and the dead pixels,over-hot pixels and blinking pixels are classified.The algorithm is applied to the actual blind-pixel detection of cooled IRFPA,and the result proves its validity.