DTCNN的人脸识别算法的Map-Reduce并行化实现研究

传统人脸识别算法都采用基于特征提取的解决方案，所以有效的特征需要很强的先验知识和丰富的工程经验.本文引入深度平 铺卷积神经网络（deep tiled convolutional neural networks，DTCNN），利用深度平铺卷积神经网络的特征学习能力来实现 人脸识别，可是由于深度平铺卷积神经网络的运算复杂度高，并且在处理海量数据时会出现训练时间过长，内存占用大等问题.为此本 文提出一种Map-Reduce并行化的DTCNN算法.实验表明，深度平铺卷积神经网络能够获得比传统经典人脸识别更好的性能，而 Map-Reduce的引入又极大地减少了大数据集下的系统训练时间.

Research of face recognition algorithm using the deep tiled convolutional neural networks and Map-Reduce method

The traditional face recognition methods were based on the feature extraction, but, the process of feature extraction needs the strong prior knowledge and plenty engineering experience. This paper introduces the deep tiled convolutional neural networks (DTCNN), which can learn the feature, to implement the face recognition, but the DTCNN will encounter many problems, such as costing too much time in training and occupying too much internal memory. This paper proposes a parallel deep tiled CNN using the framework of Map-Reduce to overcome these problems. The experimental results show that the performance of face recognition of our method is better than that of traditional method based on feature extraction, and system training time cost has been greatly decreased because of the parallel framework of Map-Reduce when testing the large scale dataset.