Author Login Editor-in-Chief Peer Review Editor Work Office Work

Computer Engineering ›› 2020, Vol. 46 ›› Issue (9): 261-267. doi: 10.19678/j.issn.1000-3428.0055895

• Graphics and Image Processing • Previous Articles     Next Articles

SKASNet: Lightweight Convolutional Neural Network for Semantic Segmentation

TAN Lei, SUN Huaijiang   

  1. School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China
  • Received:2019-09-03 Revised:2019-10-14 Published:2019-10-21

SKASNet:用于语义分割的轻量级卷积神经网络

谭镭, 孙怀江   

  1. 南京理工大学 计算机科学与工程学院, 南京 210094
  • 作者简介:谭镭(1994-),男,硕士,主研方向为图像语义分割;孙怀江,教授、博士、博士生导师。
  • 基金资助:
    国家自然科学基金(61772272)。

Abstract: Most of existing semantic segmentation models apply receptive fields of a single size in each convolution layer,which deters the models from extracting multi-scale features.To address the problem,this paper implements selective kernel convolution to build a novel residual module,Selective-Kernel-Array-Shuffle(SKAS).The selective kernel convolution can obtain the multi-scale information by adjusting the size of the receptive field.Also,a layer-wise grouped convolution method is proposed to build a lightweight network structure,SKASNet.The number of groups vary in continuous SKAS blocks in order to reduce the number of network parameters in a relatively smooth way and enhance the exchanges of information between different groups.Experimental results on the Cityscapes dataset show that the proposed network model has only 1.7 M parameters,and the segmentation accuracy reaches 68.5%.Compared with SegNet,ICNet,PSPNet and other models,the proposed model can achieve excellent segmentation performance while the number of network parameters is greatly reduced.

Key words: Convolutional Neural Network(CNN), semantic segmentation, selective kernel convolution, layer-wise grouped convolution, lightweight network model

摘要: 多数语义分割模型中的每个卷积层仅采用单一大小的感受野,不利于模型提取多尺度特征。为此,使用选择核卷积构建一个新的残差模块SKAS,通过调节感受野的大小获得多尺度信息。同时,提出一种逐层分组卷积并构建轻量级网络结构SKASNet,在连续的SKAS模块中分别使用不同的分组数,从而以相对平滑的方式降低网络参数量并增强不同分组之间的信息交流。在Cityscapes数据集上的实验结果表明,该网络模型仅有1.7 M的参数量,分割精度达到68.5%,与SegNet、ICNet和PSPNet等模型相比,其能够在大幅降低网络参数量的同时取得良好的分割效果。

关键词: 卷积神经网络, 语义分割, 选择核卷积, 逐层分组卷积, 轻量级网络模型

CLC Number: