Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (21): 250-257.DOI: 10.3778/j.issn.1002-8331.2103-0586

• Graphics and Image Processing • Previous Articles     Next Articles

Image Semantic Segmentation Network Fusing Multi-Scale and Boundary Optimization

LI Xin, ZHANG Hongying, LIU Hanyu   

  1. 1.School of Information Engineering, Southwest University of Science and Technology, Mianyang, Sichuan 621010, China
    2.Robot Technology Used for Special Environment Key Laboratory of Sichuan Province, Southwest University of Science and Technology, Mianyang, Sichuan 621010, China
  • Online:2022-11-01 Published:2022-11-01

融合多尺度和边界优化的图像语义分割网络

李鑫,张红英,刘汉玉   

  1. 1.西南科技大学 信息工程学院,四川 绵阳 621010
    2.西南科技大学 特殊环境机器人技术四川省重点实验室,四川 绵阳 621010

Abstract: About the problem that multiple convolution layers used on convolutional neural networks can cause the loss of small-scale targets in the image and the blurring of category boundaries. A stepped image semantic segmentation network structure based on multi-scale feature fusion and boundary optimization is proposed. The network aims to improve the accuracy of the network model, the spatial pooling pyramid module in the Deeplab V3+ network is optimized, and using the new activation function Funnel ReLU(FReLU) for the vision task to replace the original nonlinear activation function, which obtains accuracy compensation. Adding the optimization branch to build a ladder network and accurately predicting the boundaries of each category are used to improve the overall image segmentation accuracy and reduce the problem of the misrecognition within the category and the loss of small-scale targets in the prediction results. The experimental result on the Cityscapes dataset shows that the mean intersection over the improved network union has been significantly improved.

Key words: semantic segmentation, convolutional neural network, boundary optimization, Deeplab V3+, accuracy compensation

摘要: 针对卷积神经网络在多卷积层叠加造成的图像内小尺度目标丢失和类别边界模糊问题,提出一种基于多尺度特征融合和边界优化的阶梯型图像语义分割网络结构。该网络以提升网络模型的准确率为目标,对Deeplab V3+网络中空间池化金字塔模块进行优化,使用针对视觉任务的新激活函数Funnel ReLU(FReLU)替换原有非线性激活函数获取精度补偿,增添优化分支构建阶梯型网络,通过对各类别边界的精确预测提升整体图像分割准确率,减少预测结果中类内误识别和小尺度目标丢失问题。在Cityscapes数据集上的实验结果表明,改进后的网络各类别平均交并比指标均取得明显提升。

关键词: 语义分割, 卷积神经网络, 边界优化, Deeplab V3+, 精度补偿