摘要
以车道线检测算法为例,对网络模型进行剪枝优化。首先,对模型进行正则化,初步提高模型参数的稀疏度。之后,建立以准确率变化量和模型参数变化量为目标的多目标优化函数,通过调节准确率变化权重系数和压缩权重系数得到不同目标下的最优模型。在剪枝过程中,按卷积层深度对各层设置不同的阈值系数,最后对优化后的模型进行重新训练得到最终模型。实验结果表明:准确率优先的优化模型在车道线检测数据集的准确率为92.47%,模型压缩比为42.8%;响应速度优先的优化模型在车道线检测数据集的准确率为90.75%,模型压缩比例为67.3%。根据不同的场景需求,算法能够有效地得到不同效果的剪枝优化模型。
Taking lane detection algorithm as an example,network model is optimized by pruning.Firstly,the model is regularized to improve the sparsity of model parameters.Then,a multi-objective optimization function with the variable quantity of accuracy rate and model parameters as the objective is established.The optimal model under different objectives is obtained by adjusting the accuracy rate of changing weight coefficient and compression weight coefficient.In the pruning process,different threshold coefficients are set for each layer according to the convolution layer depth.Finally,the optimized model is retrained to obtain the final model.The experimental results show that accuracy of the optimized model with priority of accuracy in the lane line detection dataset is 92.47%,and the compression ratio of the model is 42.8%;accuracy of the optimized model with priority of response speed in the lane line detection dataset is 90.75%,and the compression ratio of the model is 67.3%.According to different scene requirements,the algorithm can effectively get different effect of pruning optimization models.
作者
翁佳昊
秦永法
唐晓峰
张浩文
关栋
WENG Jiahao;QIN Yongfa;TANG Xiaofeng;ZHANG Haowen;GUAN Dong(School of Mechanical Engineering,Yangzhou University,Yangzhou 225127,China)
出处
《传感器与微系统》
CSCD
北大核心
2023年第7期125-127,131,共4页
Transducer and Microsystem Technologies
基金
国家自然科学基金资助项目(52005433)。
关键词
深度学习
车道线检测
模型剪枝
多目标优化
deep learning
lane line detection
model pruning
multi-objective optimization
