计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (2): 177-185.DOI: 10.3778/j.issn.1002-8331.2005-0420

• 模式识别与人工智能 • 上一篇    下一篇

多尺度融合增强的图像语义分割算法

TIAN Qichuan, MENG Ying   

  1. 1.北京建筑大学 电气与信息工程学院,北京 100044
    2.北京建筑大学 建筑大数据智能处理方法研究北京市重点实验室,北京 100044
  • 出版日期:2021-01-15 发布日期:2021-01-14

Image Semantic Segmentation Algorithm with Multi-scale Feature Fusion and Enhancement

田启川,孟颖   

  1. 1.School of Electrical and Information Engineering, Beijing University of Civil Engineering and Architecture Beijing, Beijing 100044 China
    2.Beijing Key Laboratory of Intelligent Processing for Building Big Data, Beijing University of Civil Engineering and Architecture, Beijing 100044 China
  • Online:2021-01-15 Published:2021-01-14

摘要:

针对现有的图像语义分割算法存在小尺度目标丢失和分割不连续的问题,提出多尺度融合增强的图像语义分割算法,该算法在DeeplabV3+网络模型的基础上,通过构建多尺度特征提取和融合增强网络提高了对小目标特征的描述能力,使网络在分割大目标的同时也能获得小目标的特征信息,从而解决了语义分割时小尺度目标丢失和分割不连续的问题。在Cityscapes数据集上实验的结果表明,改进后的算法明显提升了小目标分割精度,解决了分割不连续的问题。最后在公开数据集PASCAL VOC 2012上进一步验证了改进算法的泛化性。

关键词: 图像语义分割, DeeplabV3+, 高分辨率信息, 小目标分割

Abstract:

For the problems of small-scale target losing and discontinuous segmentation in existing image semantic segmentation, an image semantic segmentation algorithm with multi-scale feature fusion and enhancement is proposed. Based on DeeplabV3+ network, the algorithm improves the ability to describe small target features by building a multi-scale feature extraction and fusion enhancement network. The network can also obtain small target feature while segmenting large targets, so it can solve the problem of the small target losing and the discontinuous segmentation in the semantic segmentation. Experimental results on the Cityscapes dataset show that the improved algorithm significantly improves the accuracy of small target segmentation and optimizes the problem of discontinuous segmentation. Finally, the generalization of the improved algorithm is verified on the public dataset PASCAL VOC 2012.

Key words: image semantic segmentation, DeeplabV3+, high resolution characterizations, small target segmentation