摘要
以金字塔残差模块为基础,设计了一个轻量化的网络模型。使用深度可分离卷积代替普通的卷积以减少训练的参数,同时加入通道分离模块和通道混合模块来改变特征图通道维数,以加强特征的融合。为保证网络仍然能更完整的提取特征,只对恒等映射部分进行了通道分离处理,在最后的特征融合加入了通道混合模块,在标准MPII数据集上进行测试。结果表明,轻量化的金字塔残差网络有效地减少了网络的参数,减少了约1/2的参数存储空间并保持相当的准确度,同时复杂度仅为2.83 GFLOs。
A lightweight module based on the pyramid residual module is proposed.The convolution mode is replaced by depth-wise separable convolutions structure to reduce the training parameters.Meanwhile,for enhancing the fusion of features,the channel split module and channel shuffle module are added to change the feature graph dimension.In order to make the network fully extract features,the channel split module is added on identity mapping and the channel shuffle module is added on the final features fusion.Experimental results on standard benchmarks MPII dataset show that the lightweight pyramid residual network reduces the parameters effectively.About half of the training storage space is reduced,the accuracy is comparable and the complexity is only 2.83 GFLOs.
作者
高丙坤
马克
毕洪波
王玲
GAO Bingkun;MA Ke;BI Hongbo;WANG Ling(College of Electrical and Information Engineering,Northeast Petroleum University,Daqing 163318,Heilongjiang,China)
出处
《实验室研究与探索》
CAS
北大核心
2020年第1期79-82,共4页
Research and Exploration In Laboratory
基金
东北石油大学自然科学基金项目(2017PY ZL_05,JY CX_CX06_2018和JY CX_JG06_2018)。
关键词
姿态估计
沙漏网络
轻量化
金字塔残差模块
pose estimation
hourglass network
lightweight
pyramid residual module
