基于通道剪枝的轻量化空气质量检测方法
Lightweight Air Quality Detection Method Based on Channel Pruning
崔雅博 1窦小楠 2王昆 1刘丽娜3
作者信息
- 1. 开封大学信息工程学院
- 2. 河南省地理信息院
- 3. 郑州大学河南省超算中心
- 折叠
摘要
针对传统空气质量检测系统结构复杂、部署困难以及成本较高的问题,利用图卷积网络对大气图像特征进行分析,提出了一种基于通道剪枝的轻量化空气质量检测算法.首先以ResNet50 为基础网络训练一个PM2.5 指数检测网络,实现了空气质量初步的自动化检测.然后对网络模型中的所有卷积核通道和相关的参数传递进行图节点核权重边建模,以图表示形式输入GCN,并输出针对每个卷积核节点的剪枝重要性判别预测.最后根据GCN结果进行通道剪枝,使用原始数据集对剪枝后模型的参数进行微调,在保持网络检测精准度的情况下,实现网络模型的轻量化.通过对比实验和消融实验验证了提出的检测方法具有较高的检测精度,平均检测误差仅有 5.31%,RMSE提升了 0.52,R-square仅降低了0.018,解决了网络模型的参数量和计算量过大的问题,网络参数量从 4.12×107 降低至 2.01×107,FPS从16.78 帧/s提升至30.9 帧/s,为在便携式终端上实现空气质量检测任务提供了有力的技术支持.
Abstract
Aiming at the problem of complex structure,difficult deployment and high cost of traditional air quality detection system,a lightweight air quality detection algorithm based on channel pruning was proposed using graph convolutional network to analyze atmospheric image features.Firstly,a PM2.5 index detection network was trained based on ResNet50 to achieve prelimina-ry automated detection of air quality.Then,all convolutional kernel channels and related parameter transfers in the network model were used for graph node kernel weight edge modeling,which input the graph convolutional network in the form of a graph repre-sentation and output pruning importance discrimination predictions for each convolutional kernel node.Finally,the channel prun-ing was performed according to the GCN results,and the original dataset was used to fine tune the parameters of the model after pruning,which achieved lightweight network models while maintaining network detection accuracy.The ablation experiments and comparative experiments verify that the proposed detection method has higher detection accuracy,the average detection error is only 5.31%,the RMSE is improved by 0.52,while R-square is only reduced by 0.018,which also solves the problem of exce-ssive parameter and computational complexity in the network model.The network parameter quantity is decreased from 4.12×107 to 2.01×107,and the FPS increased from 16.78 frame/s to 30.9 frame/s,which provides strong technical support for achieving air quality detection tasks on portable terminals.
关键词
空气质量检测/大气图像/通道剪枝/卷积核通道/图卷积网络/网络轻量化Key words
air quality monitoring/atmospheric image/channel pruning/convolution kernel channel/graph convolutional network/network lightweight引用本文复制引用
基金项目
河南省科技攻关项目(232102210008)
开封市科技发展计划项目(2303067)
出版年
2024
NETL
NSTL
维普
万方数据