中文

基于自适应类别权重的双重焦点损失以改进语义分割

计算机视觉与模式识别 2020-11-30 v3 机器学习

摘要

本文提出一种双重焦点损失(DFL)函数,作为标准交叉熵(CE)函数的替代,以更好地处理数据集中类别不平衡问题。我们的 DFL 方法是对近期提出的焦点损失(FL)交叉熵函数的改进,FL 提出一种缩放方法,将更大权重置于难以分类的样本而非易分类样本上。然而,FL 的缩放参数为经验设定,依赖于具体问题。此外,与其他 CE 变体一样,FL 仅关注真实类别的损失,因此无法从错误类别获得损失反馈。尽管仅关注真实样本会因 softmax 函数的特性提高真实类概率并相应降低错误类概率,但由于回避了错误类的损失,它不能实现最佳收敛。我们的 DFL 方法在两方面改进简单 FL:首先,它采纳 FL 关注难例甚于易例的思想,但以同等重要性在真实类与负类上评估损失;其次,DFL 的缩放参数被设为可学习,从而可通过反向传播自行调节,而非依赖手动调参。由此,我们提出的 DFL 方法提供了一种可自动调节的损失函数,既能减弱类别不平衡影响,又能同时更关注真实难例与负类易例。

关键词

引用

@article{arxiv.1909.11932,
  title  = {Adaptive Class Weight based Dual Focal Loss for Improved Semantic Segmentation},
  author = {Md Sazzad Hossain and Andrew P Paplinski and John M Betts},
  journal= {arXiv preprint arXiv:1909.11932},
  year   = {2020}
}

备注

We, the authors, are withdrawing this preprint due to a number of errors pointed out by the reviewers. Based on the reviewers' feedback, the paper has gone through an extensive revision, which significantly differs from this preprint version by methodologically as well as experimentally. We acknowledge the reviewers for their scrutinized review which guided our study in the right direction