增长式卷积神经网络及其在人脸检测中的应用_顾佳玲.pdfVIP

第21 卷第8 期 系 统 仿 真 学 报© Vol. 21 No. 8 2009 年4 月 Journal of System Simulation Apr., 2009 增长式卷积神经网络及其在人脸检测中的应用 顾佳玲，彭宏京 (南京工业大学信息科学与工程学院, 南京 210009) 摘 要：卷积神经网络擅于提取具有类别分辨能力的隐式特征，在人脸检测等领域获得巨大成功。 然而典型的卷积神经网络的固定结构又使得网络规模初始设定只能是经验性的，难以实现后继的再 学习。鉴于此，提出一种结构可变的卷积神经网络的构造方法：从每层只有单个卷积核的简单网 络结构开始训练，逐渐为各网络层增加新的卷积神经元并修改新增连接权重，当训练结果达到预 期目标时训练结束。在人 脸检测的实验中，不同网络规模下的识别结果表明基于结构增长生成的 网络可以在精确度和网络规模之间取得一个非常好的折衷。此外，在追加新的学习样本时，分类 器在保持原有学习结果基础上，只需调整少量新增神经元的权值，就能明显提高检测率。 关键词：卷积神经网络；机器学习；人脸检测；神经网络；模式分类 中图分类号：TP391 文献标识码：A 文章编号：1004-731X (2009) 08-2441-05 Incremental Convolution Neural Network and Its Application in Face Detection GU Jia-ling, PENG Hong-jing (The College of Information Science and Engineering, Nanjing University of Technology, Nanjing 210009, China) Abstract: Convolution neural networks have the capability of extracting hidden discriminating features, so that they successfully have been applied in the fields of face detection, etc. Due to their fixed architecture, the scale of neural networks usually depends on the designer’s experience and prevents them from learning further. A novel approach of constructing CNN with varying architecture was proposed. Training neural networks started off with the simplest architecture with single neuron each layer, then adding new neural cells in every layer and modifying their corresponding weights continuously, until training target was reached. In the face detection