Abstract: Aimed at the problem that the face recognition accuracy is low under unconstrained condition, a granular neural network and genetic optimization algorithm based face recognition algorithm is proposed. Firstly, the face dataset is initially analyzed to decide the faces distribution in each granularity, and the data points with the same complexity level are grouped to a group; then, the faces are divided to front, eyes, mouth, and the sub-regions of the face are trained by granular neural network with different counts of data points, and multi training results are generated; lastly, a multi-level genetic algorithm is designed to optimize the granular neural network. Two public face databases based compared experimental results show that the proposed algorithm releases better face recognition accuracy than the other algorithms.

Key words: face recognition, genetic algorithm, neural networks, data complexity, granular selection

摘要： 针对非限条件下人脸识别准确率较低的问题，提出一种基于粒神经网络（MNN）与遗传算法优化的人脸识别算法。对人脸库进行初始化分析决定每个粒子中人脸的分布，将同一复杂度级别的数据分为一组；将人脸分为额头、眼睛与嘴三个部分，粒神经网络采用不同数量的数据点对面部子区域进行训练，获得多个训练结果；设计了一种多级的遗传算法对粒神经网络进行优化。基于两组公开人脸数据库的对比实验结果表明，该算法的识别准确率优于其他人脸识别算法。

关键词: 人脸识别, 遗传算法, 神经网络, 数据复杂度, 粒子选择