A number of current face recognition algorithms use whole face representations found by statistical methods. Independent Component Analysis(ICA) is an example of such methods which is based on signal high-order statistic characteristics. While such unavoidable external factors as illumination

posture and information deformity will cause great changes of gray-scale image data

and eventually will decrease the stability of recognition. This paper presents a local face recognition algorithm that is based on ICA and the nearest feature line (NFL). Firstly

by using manually aligned eye position

segmenting a face image into two parts according to the geometric characteristics of human face

removing hair style and other useless information

then processing principal component analysis (PCA) and ICA for respective parts

and calculating corresponding NFL distance

ultimately processing comprehensive recognition by setting reasonable coefficient of weight. Compared with traditional holistic image representation

this method has many advantages

such as a much higher recognition rate

more stable and flexible in practice. Through a number of experiments

it proves to be an efficient human face recognition algorithm.