Current Issue Cover
  • 发布时间:
  • 摘要点击次数:
  • 全文下载次数:
  • DOI:
  • 2020 | Volume  | Number 6

()

摘 要
目的 图像语义分割是指通过识别每个像素的语义类别,将输入图像按照语义类别分割成不同的区域。在自动驾驶、视频监控和刑侦分析等领域有重要应用价值。然而传统图像语义分割需要的像素级标注数据难以大量获取,因此图像语义分割的弱监督学习是当前的重要研究方向。弱监督学习是指使用弱标注样本完成监督学习,弱标注比像素级标注的标注速度快、标注方式容易,包括散点、边界框、涂鸦等标注方式。方法 针对现有方法对多层特征利用不充分的问题,提出了一种基于动态掩膜生成的弱监督语义分割方法。该方法以边界框作为初始前景分割轮廓,使用迭代方式利用卷积神经网络(CNN) 多层特征获取前景目标的边缘信息,根据边缘信息生成掩膜。迭代的过程中首先使用高层特征对前景目标的大体形状和位置做出估计,得到粗略的物体分割掩膜。然后根据已获得的粗略掩膜,逐层使用CNN 特征对掩膜做更新。结果 在Pascal VOC 2012 数据集上取得了78.06% 的分割精度,相比现有弱监督语义分割方法,边界框监督、弱-半监督、掩膜排序和实例剪切方法,分别提高了14.71%、4.04%、3.10% 和0.92%。结论 该方法能够利用高层语义特征,减少分割掩膜中语义级别的错误;同时使用底层特征对掩膜做更新可以提高分割边缘的准确性。
关键词

()

Abstract
Objective Image semantic segmentation refers to segmenting input images into different regions according to semantic categories by identifying the semantic category of each pixel. Traditional semantic segmentation methods need a lot of pixel-level annotation data and thus, weakly supervised learning is getting more attention. Weakly supervised learning makes use of weak label which is faster and easier to get, such as points, bounding box, scribble, etc, for training. Method To solve the problem of missing edge information in weak label for semantic segmentation, a weakly supervised semantic segmentation method based on dynamic mask generation is proposed. We first use the bounding box as initial foreground edge contour, and then iteratively adjust it with multi-layer features of convolutional neural networks, and finally generate masks from it. Result The segmentation accuracy of our method on Pascal VOC 2012 dataset is 78.06%. Compared with current weakly supervised semantic segmentation methods, Boxsup、WSSL、SDI and CaP, the accuracy increases by 14.71%, 4.04%, 3.10% and 0.92% respectively. Conclusion Highlevel features are used for estimating the approximate shape and position of the foreground object and generate rough edges, which would be corrected layer by layer with multi-layer features. High-level semantic features could decrease edge information error in semantic level and low-level image features improve the accuracy of the edge.
Keywords
QQ在线


订阅号|日报