指数弹性动量卷积神经网络及其在行人检测中的应用

针对深度卷积神经网络存在规则化参数多、未利用浅层先验知识、参数随机初始化后易导致权值更新梯度弥散及训练早熟等问题,采用PCA非监督学习方式获取导向性初始化参数数值方法,并基于对网络误差的传播分析,提出指数自适应弹性动量参数学习方法.以复杂场景下行人目标为例进行目标检测试验,实验表明:与人工特征检测识别方案及传统深度卷积模型相比,该模型可有效提升目标检测精度,检测速度提升20%以上;与其他动量同源更新机制相比,该算法收敛速度更快,收敛曲线更平滑,泛化能力强,可在不同深度模型均可取得较好检测效果,准确率分别平均提高1.6%,1.8%和6.19%.

A deep convolution neural network for object detection based

convolutional neural network (CNN) has too many parameters to initialize, and the usual random initialization method is easy to disappear of modified gradient and the problem of premature. The unsupervised PCA learning method is used to obtain oriented initialization parameters. And the gradient descendent method with exponential flexible momentum for updating free parameters of the network is proposed on the basis of analyzing the error propagation of the network. Image detection experiments are respectively carried out on pedestrian detection, and the results show that, compared with other artificial feature detection algorithms, this method can effectively improve target detection accuracy and the detection speed of this method is 20% faster than that of classical CNN; compared with homologous updating mechanism of other momentum, our method has faster convergence and smaller oscillation, and can improve the detection accuracy by 1.6%, 1.8% and 6.19% respectively in different depth models.