基于组注意力卷积和特征重建的轻量分割算法  

Lightweight segmentation algorithm based on group attention convolution and feature reconstruction

作　　者：胡国清[1] 谭海亮 戈明亮[1] HU Guo-qing;TAN Hai-liang;GE Ming-liang(School of Mechanical and Automotive Engineering,South China University of Technology,Guangzhou 510641,China)

机构地区：[1]华南理工大学机械与汽车工程学院,广东广州510641

出　　处：《计算机工程与设计》2022年第1期120-126,共7页Computer Engineering and Design

基　　金：广东省自然科学基金项目(2016A030313520)。

摘　　要：针对卷积神经网络的庞大参数量和计算量难以应用于移动设备或嵌入式设备的问题,提出冗余特征重建模块(redundant feature reconstruction,RFR)和组注意力卷积模块(group attention convolution,GAC),RFR模块使用较少的参数量提取重要的固有特征,通过线性算子重建冗余特征帮助网络模型推理,GAC模块通过引入注意力机制改善分组卷积中各分组之间的信息交互不充分的问题,基于所提出模块构建一种高效轻量化的语义分割网络。在Cityscapes数据集上进行的实验显示了所提网络模型达到了良好的分割准确率,参数量和网络模型大小低于当前典型语义分割算法,在算法复杂度和性能上达到了平衡,表明了其有效性。Aiming at the problem that the huge amount of parameters and calculations of deep convolutional neural networks are difficult to apply to mobile devices or embedded devices,the redundant feature reconstruction(RFR)module and group attention convolution(GAC)module were proposed.The RFR module used a small amount of parameters to extract important inherent features,and used linear operators to reconstruct redundant features to help network model reasoning.The GAC module improved the problem of insufficient information interaction between groups of group convolution by introducing attention mechanism,and an efficient and lightweight semantic segmentation network was constructed based on the proposed module.Experiments on Cityscapes dataset show that the proposed network model achieves good accuracy,while the parameter amount and network model size are lower than current typical semantic segmentation algorithms,achieving a balance between algorithm complexity and performance,indicating the effectiveness of the proposed module.

关 键 词：卷积神经网络 特征重建 注意力机制 组卷积 轻量化 语义分割 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]