计算机工程 ›› 2020, Vol. 46 ›› Issue (5): 254-258,266.doi: 10.19678/j.issn.1000-3428.0054964

• 图形图像处理 • 上一篇    下一篇

基于增强特征融合解码器的语义分割算法

马震环1,2, 高洪举3, 雷涛1   

  1. 1. 中国科学院光电技术研究所, 成都 610209;
    2. 中国科学院大学 电子电气与通信工程学院, 北京 100049;
    3. 32183部队, 辽宁 锦州 121000
  • 收稿日期:2019-05-20 修回日期:2019-07-17 发布日期:2019-08-06
  • 作者简介:马震环(1994-),男,硕士研究生,主研方向为图像语义分割、深度学习;高洪举,高级工程师;雷涛,研究员。
  • 基金项目:
    中国科学院青年创新促进会基金(2016336)。

Semantic Segmentation Algorithm Based on Enhanced Feature Fusion Decoder

MA Zhenhuan1,2, GAO Hongju3, LEI Tao1   

  1. 1. Institute of Optics and Electronics, Chinese Academy of Sciences, Chengdu 610209, China;
    2. School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, Beijing 100049, China;
    3. 32183 Troops, Jinzhou, Liaoning 121000, China
  • Received:2019-05-20 Revised:2019-07-17 Published:2019-08-06

摘要: 针对语义分割中全卷积神经网络解码器部分特征融合低效的问题,设计一种增强特征融合的解码器。级联深层特征与降维后的浅层特征,经过卷积运算后引入自身平方项的注意力机制,通过卷积预测自身项与自身平方项各通道的权重,利用乘法增强后对结果进行作和。基于pascal voc2012数据集的实验结果表明,该解码器相比原网络mIoU指标提升2.14%,结合不同特征融合方式的解码结果也验证了其性能优于同一框架下的其他对比方法。

关键词: 语义分割, 卷积神经网络, 解码器, 特征融合, 注意力机制

Abstract: To address the inefficient fusion of part of features in full Convolutional Neural Network(CNN) decoder in semantic segmentation,this paper proposes an enhanced feature fusion decoder.The decoder cascades high-level features and low-level features after dimensionality reduction.Then,after the convolution operations,it introduces the attention mechanism of its squared term,and predicts the weights of each channel of its own term and its squared term by convolution.Finally,the weights are enhanced by multiplication and added to get a sum.Experimental results on the pascal voc2012 dataset show that,compared with the original network,the proposed method increases the value of mIoU index by 2.14%.Decoding results under different ways of feature fusion also demonstrates that it outperforms other methods under the same framework.

Key words: semantic segmentation, Convolutional Neural Network(CNN), decoder, feature fusion, attention mechanism

中图分类号: