摘要
针对现有部件分割精度较低、泛化性和精度无法兼顾等问题,文中提出基于DeepLab的物体部件分割网络(DeepLab-MAFE-DSC).网络的编码器部分提出多尺度自适应形态特征提取模块(MAFE),利用可形变卷积增强模型对不规则轮廓的处理能力,并采取先级联再并行相加的采样模式,兼顾全局和局部信息.解码器部分设计基于跳跃式架构的解码器模块(DSC),同时连接深层的语义信息和浅层的表征信息.在数据集上的实验表明,DeepLab-MAFE-DSC具有简单、分割精度较高、泛化性较强的优点.
The low precision exists in the existing part segmentation,and the generalization and precision can not be balanced.Aiming at the problems,a part segmentation network(DeepLab-MAFE-DSC)based on DeepLab is proposed.A multi-scale adaptive-pattern feature extraction(MAFE)module is proposed in encoder part of the network.The deformable convolution is exploited to enhance the processing capability to irregular contour,and sampling mode of cascade and concatenate in parallel is adopted to balance global and local information simultaneously.A decoder module based on skip connection(DSC)is designed to connect high-level semantic information and low-level character information.Experiments on the dataset show the advantages of DeepLab-MAFE-DSC in simplicity,high part segmentation accuracy and strong generalization.
作者
赵霞
倪颖婷
ZHAO Xia;NI Yingting(College of Electronics and Information Engineering,Tongji University,Shanghai 201804)
出处
《模式识别与人工智能》
EI
CSCD
北大核心
2020年第3期211-220,共10页
Pattern Recognition and Artificial Intelligence
基金
上海航天科技创新基金项目(No.SAST2016018)资助。
关键词
卷积神经网络
物体部件分割
可形变卷积
Convolutional Neural Network
Object Part Segmentation
Deformable Convolution