多层特征融合的PCANet及其在人脸识别中的应用

Multi-stage Feature Fusion PCANet and Its Application to Face Recognition

主成分分析网络(PCANet)是一种简单的深度学习网络模型,在图像识别领域具有很强的应用潜力.本文在PCANet的基础上,通过对PCANet结构进行分析,构造了一种基于多层特征融合的PCANet(PCANet_dense)网络模型.与单纯地只将前一层网络输出作为后一层网络输入的PCANet不同,PCANet_dense利用了不同层的特征信息.在2层网络结构中,它首先将原始图像特征和第1层网络的输出进行级联,然后将级联后的结果作为第2层网络的输入.而在3层网络结构中,它则将第1层和第2层网络的输出级联起来,作为第3层网络的输入.由于PCANet_dense在训练每一层(除了第1层)时使用了更多信息,因此能够获得比原PCANet更好的效果.为了验证所提方法的有效性,本文使用CMU PIE数据集构建网络模型,并在ORL、AR和Extended Yale B 3个公开人脸数据集上对所提出方法的性能进行了测试,实验结果表明,本文提出的PCANet_dense获得了比PCANet更好的性能.

Principal component analysis network(PCANet)is a simple deep learning network model but shows strong application potential in many fields,especially in image recognition.In this paper,a new multi-stage feature fusion model based PCANet(PCANet_dense)is constructed by analyzing the structure of PCANet.Different from PCANet which only takes the output of the former layer as the input of the latter layer,PCANet_dense uses the feature information of different layers.In the two-layer network structure,it firstly cascades the original image features with the output of the first layer network,and then takes the fused features as the input of the second layer network.In the three-layer network structure,it cascades the output of the first and the second layer in series as the input of the third layer network.Due to the use of more information in each layer(except the first layer)in training process,the new model gets good performance.In order to verify the effectiveness of the proposed method,several comparative experiments on different face datasets,such as ORL,AR and Extended Yale B,are established.The network model of PCANet and PCANet_dense are well constructed on CMU PIE dataset before all the experiments.The experimental results show that the proposed multi-stage feature fusion PCANet method achieves better performance than PCANet.