作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2021, Vol. 47 ›› Issue (4): 256-261. doi: 10.19678/j.issn.1000-3428.0058015

• 图形图像处理 • 上一篇    下一篇

基于改进SEGNET模型的图像语义分割

罗嗣卿, 张志超, 岳琪   

  1. 东北林业大学 信息与计算机工程学院, 哈尔滨 150040
  • 收稿日期:2020-04-09 修回日期:2020-05-21 发布日期:2020-05-28
  • 作者简介:罗嗣卿(1964-),男,副教授、硕士,主研方向为图像处理、信息系统集成;张志超、硕士研究生;岳琪(通信作者),教授、博士。
  • 基金资助:
    国家自然科学基金青年项目“基于用户标签和主体兴趣的社会媒体信息推荐研究”(61806049)。

Semantic Image Segmentation Based on Improved SEGNET Model

LUO Siqing, ZHANG Zhichao, YUE Qi   

  1. School of Information and Computer Engineering, Northeast Forestry University, Harbin 150040, China
  • Received:2020-04-09 Revised:2020-05-21 Published:2020-05-28

摘要: 使用原始SEGNET模型对图像进行语义分割时,未对图像中相邻像素点间的关系进行考虑,导致同一目标中像素点类别预测结果不一致。通过在SEGNET结构中加入一条自上而下的通道,使得SEGNET包含的多尺度语义信息更加丰富,从而提升对每个像素点的类别预测精度,在模型中加入生成对抗网络以充分考虑空间中相邻像素点间关系。实验结果表明,该模型的语义分割效果相比原始SEGNET模型显著提升,且可有效解决SEGNET测试中出现的分类错误问题。

关键词: SEGNET模型, 生成对抗网络, 多尺度语义信息, 相邻像素类别关系, 特征融合

Abstract: When applied to semantic image segmentation,the original SEGNET model does not account for the relationship between adjacent pixels in the image,resulting in inconsistent prediction results of pixel categories in the same target.By adding a top-down channel in the SEGNET structure,the multi-scale semantic information of the SEGNET model is enriched,and the accuracy of category prediction for each pixel is improved.The generative adversarial network is added to the model to ensure that the model can consider the relationship between adjacent pixels in space.The experimental results show that the semantic segmentation effect of the improved SEGNET model is significantly improved compared with the original SEGNET model.It can effectively avoid the classification errors in the SEGNET test.

Key words: SEGNET model, Generative Adversarial Network (GAN), multi-scale semantic information, adjacent pixel category relationship, feature fusion

中图分类号: