PSdropout卷积神经网络在危化品巡检车中的应用

PDF下载HTML

Title:

Application of PSdropout convolutional neural network in inspection car for hazardous chemicals

作者:

沙芸, 李齐飞, 甘建旺, 刘学君, 隗立昂

北京石油化工学院 信息工程学院, 北京 102617

Author(s):

SHA Yun, LI Qifei, GAN Jianwang, LIU Xuejun, WEI Li’ang

School of Information Engineering, Beijing Institute of Petrochemical Technology, Beijing 102617, China

关键词:

危化品仓储; 巡检车; 卷积神经网络; 神经元筛选; 泊松分布; 子网络; 全链接层; 网络架构

Keywords:

storage of hazardous chemicals; inspection vehicle; convolutional neural network; dropout; Poisson distribution; subnetwork; full link layer; network architecture

分类号:

TP391.4;TQ086.5

DOI:

10.11992/tis.202007022

摘要:

危化品仓储环境复杂多变，基于卷积神经网络的视觉巡检车需要快速的训练方法以便适用不同的环境，提高卷积神经网络的训练速度是当前亟待解决的问题。迅速在网络中提取有效的神经元，是提高算法训练速度的关键。传统的算法中，全链接层神经元的去留问题通常采用基于伯努力分布假设的Dropout方法，本文提出一种基于泊松分布的Dropout方法。理论上看，在充分利用神经元历史行为的基础上，基于泊松分布与基于伯努力分布的最大似然函数类似。实验结果表明，在保持正确率的情况下，训练提前收敛，节约了训练时间。

Abstract:

The storage environment of hazardous chemicals is complex and changeable. To adapt to different environments, visual-inspection vehicles based on convolutional neural networks require fast training methods, but the problem of improving the training speed of convolutional neural networks remains to be solved. To improve the training speed of the algorithm, effective neural cell in network must be extracted from the network more quickly. In the traditional algorithm, the removal of neurons from the fully connected layer is typically performed using the dropout method, which is based on the Bernoulli distribution hypothesis. In this paper, we propose a dropout method based on the Poisson distribution. Theoretically, by making full use of the historical behavior of neurons, the maximum likelihood function based on the Poisson distribution is similar to that based on the Bernoulli distribution. The experimental results show that while maintaining the correct rate, the training can be converged in advance, thereby saving the training time.