摘要
针对基于卷积神经网络的遥感目标检测模型压缩问题,基于卷积核剪枝理论,设计了卷积通道剪枝的方案,对YOLOv3模型进行精简压缩。提出卷积通道的概念并将BN层系数作为卷积通道的评估因子,使用L1正则化将评估因子稀疏化,将评估因子值较小的卷积通道剔除,再对各卷积层中的参数进行微调,从而达到模型压缩的目的。使用该方法对自制的典型遥感目标检测数据集进行目标检测实验,在剔除90%参数的情况下,测试精度下降率在10%以内。实验结果表明该方法能以较小的精度损失为代价,节省大部分的储存空间和计算量。
Considering the compression problems of remote sensing object detection model based on convolutional neural network,a convolution channel pruning scheme based on convolution kernel pruning theory is designed,and compressed the YOLOv3 model is simply compressed.The concept of convolution channel is proposed and the coefficient in BN layer is used as the assessment factor of the convolution channel.The assessment factor is sparsed by L1 regularization,the convolution channel with smaller fine tuning assessment factor value is eliminated,and the parameters in each convolution layer are.To achieve the purpose of model compression.This method is used to test the self-made typical remote sensing object detection dataset.Under the condition of 90%of parameters removed,the decline rate of test accuracy is within 10%,the experiment results indicate that the method save most parts of storage space and calculation amount at the cost of less precision.
作者
韩要昌
王洁
鲁力
李宇环
HAN Yao-chang;WANG Jie;LU Li;LI Yu-huan(Air and Missile Denfense College,Air Force Engineering University,Xi’an 710051,China)
出处
《火力与指挥控制》
CSCD
北大核心
2021年第2期23-29,共7页
Fire Control & Command Control
基金
国家自然科学基金资助项目(61601499)。
关键词
卷积神经网络
遥感目标检测
YOLOv3
模型压缩
convolutional neural network
remote sensing object detection
YOLOv3
model compression
