摘要
针对复杂卷积神经网络(CNN)在中小型人脸数据库中的识别结果容易出现过拟合现象,提出一种基于改进CNN网络与集成学习的人脸识别算法。改进CNN网络结合平面网络和残差网络的特点,采用平均池化层代替全连接层,使得网络结构简单且可移植性强。在改进CNN网络的基础上,利用基于投票法的集成学习策略将所有个体学习器结果凸组合为最终结果,实现更准确的人脸识别。实验结果表明,该算法在Color FERET、AR和ORL人脸数据库上的识别准确率分别达到98.89%、99.67%和100%,并且具有较快的收敛速度。
To address overfitting of recognition results in complex Convolutional Neural Network(CNN)on small and medium face databases,this paper proposes a face recognition algorithm based on improved CNN and ensemble learning.Combining the characteristics of planar networks and residual networks,the improved CNN replaces the fully connected layer with the average pooling layer to make the network structure simple and highly portable.Based on this improved CNN,the voting-based ensemble learning strategy is used to implement convex combination for results of all individual learners and obtain the final result,so more accurate face recognition could be realized.Experimental results show that the recognition accuracy of the proposed algorithm reaches 98.89%,99.67%and 100%respectively on Color FERET,AR and ORL face databases with a high convergence speed.
作者
柯鹏飞
蔡茂国
吴涛
KE Pengfei;CAI Maoguo;WU Tao(College of Information Engineering,Shenzhen University,Shenzhen,Guangdong 518060,China)
出处
《计算机工程》
CAS
CSCD
北大核心
2020年第2期262-267,273,共7页
Computer Engineering
基金
国家自然科学基金(61872244)
关键词
深度学习
模式识别
卷积神经网络
集成学习
人脸识别
deep learning
pattern recognition
Convolutional Neural Network(CNN)
ensemble learning
face recognition
