优化非线性激活函数-全局卷积神经网络的物体识别算法

Object Recognition Algorithm Based on Optimized Nonlinear Activation Function-global Convolutional Neural Network

传统的物体识别算法识别精度、自适应能力弱等问题已然不能满足实际的仓储物流领域对物体识别精度的要求.近年来,相关学者提出了基于深度学习的物体识别算法,它得到一定的推广和应用.但是,深度学习在物体识别的应用过程中存在以下问题:一是深度学习模型中激活函数的非线性建模能力弱;二是深度学习模型大量重复的池化操作丢失信息.鉴于此,本文提出了一种参数形式统一且可学习的指数非线性单元(Multiple Parameters Exponential Linear Units,MPELU).它通过在ELU(Exponential Linear Units)中引入两个学习的参数,提升模型的非线性建模能力.同时,本文提出了一种新的全局卷积神经网络结构,减少大量池化操作丢失特征信息的问题.基于上述思想,本文提出了优化非线性激活函数-全局卷积神经网络的物体识别算法.利用本文所提算法对CIFAR100数据集和ImageNet数据集分别进行实验.结果表明,本文所提物体识别方法不仅识别准确率较传统机器学习、其他深度学习模型有较大幅度提升,而且具有良好的稳定性和鲁棒性.

The problems of traditional object recognition algorithms such as recognition accuracy and weak self-adaptive ability cannot meet the requirements of the actual warehouse logistics field for object recognition accuracy.Related scholars have proposed an object recognition algorithm based on deep learning,which has been promoted and applied to some extent.However,deep learning has the following problems in the application process of object recognition:one is that the nonlinear modeling ability of the activation function in the deep learning model is weak;the other is that a large number of repeated pooling operations of the deep learning model lose information.So,this paper proposes a multiple parameters exponential linear units(MPELU).It introduces two learned parameters in exponential linear units(ELU)to improve the nonlinear modeling ability of the model.At the same time,this paper proposes a new global convolutional neural network structure to reduce the problem of losing feature information for a large number of pooling operations.Based on the above ideas,this paper proposes an object recognition algorithm that optimizes the nonlinear activation function-global convolutional neural network.The algorithm proposed in this paper was used to conduct experiments on the CIFAR100 dataset and the ImageNet dataset.The results show that the object recognition method proposed in this paper not only has a greater improvement in recognition accuracy than traditional machine learning and other deep learning models,but also has good stability and robustness.