In this paper we propose a novel data driven strategy for designing Gabor wavelets for face recognition. Each face image is represented through a multi-sensor scheme, which splits the 2D frequency plane into a number of channels and identifies the most significant units for extracting information. The representative units for a set of face images are then derived based on statistical analysis of these units. The locations of these units in the 2D frequency plane are then used to design the frequency and orientation of Gabor wavelets for face recognition. Once frequency and orientation are determined, the scale of a Gabor wavelet is determined by the sharpness of the filtered images. Two Gabor wavelet based face recognition algorithms are applied to demonstrate the advantages of the proposed strategy against conventional parameter settings. Experimental results show that the face recognition algorithms using the designed Gabor wavelets achieve better performance in terms of accuracy and efficiency. Since the strategy is based on the training data, it can be easily applied to designing Gabor wavelets for general pattern recognition task.