摘要
针对目前卷积神经网络在图像语义分割上存在运算效率的不足,考虑实际嵌入式移动设备应用中对网络模型大小、运算速度和能耗的需求,研究和讨论了语义分割网络参数的压缩方法和网络结构的优化方式,并利用深度可分离卷积、多尺寸卷积核融合和金字塔池化的方法,提出快速、准确的小尺寸语义分割网络模型。该网络模型在Cityscapes数据集上进行训练和测试,在模型尺寸、运算时间和分割精度等方面都有良好的表现,能够同时兼顾语义分割任务中对效率和精度的要求,提高了语义分割技术在嵌入式移动设备上的实用性。
In order to solve the problem of insufficient computational efficiency of current convolutional neural network on image semantic segmentation, considering the requirements of network size, computing speed and energy consumption of the embedded mobile device applications, the compression and optimization methods of semantic segmentation network are discussed. Then, an efficient semantic segmentation network is proposed by using the depth-wise separable convolution, the muhi-dimension-kernel convolution and the pyramid pooling. The proposed network is trained and tested on the Cityscape dataset and shows promising performance in net- work size, operation time and segmentation precision. It provides the practicability of semantic segmentation technology for embedded mobile devices.
作者
敖焕轩
李煊鹏
张为公
AO Huan-xuan;LI Xuan-peng;ZHANG Wei-gong(School of Instrument Science and Engineering,Southeast University,Nanjing 210096,China)
出处
《测控技术》
CSCD
2018年第11期40-44,共5页
Measurement & Control Technology
基金
江苏省基础研究计划(自然科学基金)(BK20160700)
中央高校基本科研业务费(2242018K40067)
关键词
卷积神经网络
语义分割
效率
深度可分离卷积
金字塔池化
convolutional neural network
semantic segmentation
efficiency
depth-wise separable convolution
pyramid pooling
