Abstract: In order to extract more effective classification features from the Gabor filtered maps, a novel face recognition method based on GDP（Gabor Directional Pattern） is proposed. Firstly, multi-scale and multi-orientation Gabor filters are used to extract their corresponding Gabor magnitude maps. Then, a novel operator named GDP is proposed, which converts the Gabor transformed images into several GDP maps by encoding all the Gabor magnitude maps with the same scale. Finally the spatial histograms of all the GDP maps are concatenated together to represent the facial appearances. The extracted GDP feature not only has good robustness but also has low dimensionality. Experimental results on the ORL and CAS-PEAL databases show that the performance of proposed method is superior to other Gabor based methods such as LGBP（Local Gabor Binary Pattern）and LGXP（Local Gabor XOR Pattern） with a significantly lower feature dimension.

Key words: face recognition, Gabor filter, Gabor Directional Pattern（GDP）, feature extraction

摘要： 为了从Gabor滤波后的幅值图中提取更加有效的分类特征，提出了一种新的基于Gabor定向模式（GDP）的人脸识别方法。首先对人脸图像进行多尺度多方向的Gabor滤波，然后提出了一种新的GDP算子通过对每种尺度下所有方向的Gabor幅度图进行编码得到每种尺度对应的GDP模式图，最后将所有GDP模式图的直方图向量串联作为最终的人脸表示。由于GDP算子同时对同一尺度下的所有方向上的Gabor幅度响应进行编码，因而GDP特征不仅对外界变化具有较好的鲁棒性，而且能够显著降低最终的特征长度。在ORL和CAS-PEAL人脸库上的实验结果显示GDP方法能以更小的特征长度获得优于传统LGBP及LGXP等方法的识别效果，证明了方法的有效性。

关键词: 人脸识别, Gabor滤波, Gabor定向模式, 特征提取