激光与光电子学进展, 2018, 55 (10): 101504, 网络出版: 2018-10-14
特征匹配融合结合改进卷积神经网络的人脸识别 下载: 1294次
Face Recognition by Feature Matching Fusion Combined with Improved Convolutional Neural Network
机器视觉 卷积神经网络 特征匹配 特征融合 人脸识别 machine vision convolutional neural network feature matching feature fusion face recognition
摘要
提出一种特征匹配融合结合改进卷积神经网络(CNN)的图像识别方法。针对局部二值模式(LBP)算子提取的特征信息局限以及不能准确描述图像轮廓信息的问题, 使用梯度方向直方图(HOG)和LBP分层特征融合的方法在卷积神经网络中对训练集进行特征提取, 再将匹配提取的特征图像输入改进的卷积神经网络进行训练、识别。以ReLU为激活函数, 输出层用Softmax分类器, 并通过TensorFlow框架进行训练, 在ORL、YALE和CAS-PEAL人脸库上进行人脸识别仿真, 该方法识别率分别达到了99.2%、98.7%、97.2%, 高于其他对比算法。
Abstract
This paper presents an image recognition method based on feature matching fusion and improved convolutional neural network. Aiming at the problem that the texture features extracted by the local binary pattern (LBP) descriptors are limited and cannot describe the image edge and direction information effectively, the feature extraction of the training set is performed in the convolutional neural network by the histogram of oriented gradient (HOG) and LBP hierarchical feature fusion method. Then the extracted feature pictures are input into the improved convolutional neural network for training and recognition. The simulations are performed on ORL, YALE and CAS-PEAL face databases with ReLU as the activation function and the output layer with the Softmax classifier, and trained on the TensorFlow framework. The recognition rate of the proposed method reaches 99.2%, 98.7%, and 97.2% respectively, which is higher than other algorithms for comparison.
李佳妮, 张宝华. 特征匹配融合结合改进卷积神经网络的人脸识别[J]. 激光与光电子学进展, 2018, 55(10): 101504. Li Jiani, Zhang Baohua. Face Recognition by Feature Matching Fusion Combined with Improved Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2018, 55(10): 101504.