A projection function called minimal neighborhood mean projection function (MNMPF) is proposed. The projection function calculates and stores the minimal neighborhood mean of each pixel on each projection line, so that it is able to trace the low grayscale features in image. Compared with traditional projection functions, i.e. integral projection function (IPF) and variance projection function (VPF), MNMPF is insensitive to sheet noise, due to the local selectivity of its minimum operation. During the computation of MNMPF, the image locations of minima are recorded at the same time. This makes MNMPF a 2D operator. All these properties of MNMPF are very suitable for eye location. It can bring precise and robust response to eyes, especially pupils. Experiments on CAS-PEAL and BioID databases show its excellent correct rate and precision over traditional projection functions.