计算机科学与探索 ›› 2021, Vol. 15 ›› Issue (5): 949-957.DOI: 10.3778/j.issn.1673-9418.2005055

• 图形图像 • 上一篇    下一篇

分级特征融合的图像语义分割

赵小强,徐慧萍   

  1. 1. 兰州理工大学 电气工程与信息工程学院,兰州 730050
    2. 甘肃省工业过程先进控制重点实验室,兰州 730050
    3. 兰州理工大学 国家级电气与控制工程实验教学中心,兰州 730050
  • 出版日期:2021-05-01 发布日期:2021-04-30

Image Semantic Segmentation Method with Hierarchical Feature Fusion

ZHAO Xiaoqiang, XU Huiping   

  1. 1. College of Electrical and Information Engineering, Lanzhou University of Technology, Lanzhou 730050, China
    2. Key Laboratory of Gansu Advanced Control for Industrial Processes, Lanzhou 730050, China
    3. National Experimental Teaching Center of Electrical and Control Engineering, Lanzhou University of Technology, Lanzhou 730050, China
  • Online:2021-05-01 Published:2021-04-30

摘要:

在图像语义分割中,利用卷积神经网络对图像信息进行特征提取时,针对卷积神经网络没有有效利用各层级间的特征信息而导致图像语义分割精度受损的问题,提出分级特征融合的图像语义分割方法。该方法利用卷积结构分级提取含有像素级的浅层低级特征和含有图像级的深层语义特征,进一步挖掘不同层级间的特征信息,充分获取浅层低级特征和深层语义特征中隐藏的特征信息,接着通过上采样操作细化浅层低级特征信息后对所有特征信息进行合并融合,最终实现分级特征融合的图像语义分割。在实验方面,通过多次分解实验验证了所提方法在不同阶段所提取的特征信息和不同网络深度时的特征信息对语义分割结果的影响。同时在公认的图像语义分割数据集PASCAL VOC 2012上,与3种主流方法进行实验对比,结果显示所提方法在客观评价指标和主观效果性能方面均存在优越性,从而验证了该方法可以有效地提升语义分割的精度。

关键词: 卷积神经网络(CNN), 特征提取, 特征融合, 图像语义分割

Abstract:

When image feature information extraction is performed by using convolutional networks in the image  semantic segmentation, the feature information between levels is not used effectively, resulting in image segmentation accuracy impaired. Therefore, an image semantic segmentation method of hierarchical feature fusion is proposed. The method uses the convolution structure to extract the low-level features at pixel level and the deep-level semantic features at image-level, the hidden features of different levels are delved step by step and the feature information hidden in the low-level features and deep-level semantic features are obtained fully. The upsampling is used to refine the low-level feature information and all the feature information is fused successively. Finally, the image semantic segmentation of hierarchical feature fusion is realized. In the aspect of experiment, through multiple decomposition experiments, the effect of this method is verified, the feature information extracted at different stages and the feature information extracted at different network depths, on the semantic segmentation results. At the same time, compared with three main methods on the image semantic segmentation PASCAL VOC 2012 dataset, the objective evaluation index and subjective effect performance results indicate that the proposed method is superior. It is verified that this method can effectively improve the accuracy of semantic segmentation.

Key words: convolutional neural networks (CNN), feature extraction, feature fusion, image semantic segmentation