• 主题
高级搜索  >
历史搜索 清空历史
首页> 外文会议>International Conference on Computer Vision, Image and Deep Learning >Balance-batch: An Optimized Method for Semantic Segmentation Loss Functions
【24h】

Balance-batch: An Optimized Method for Semantic Segmentation Loss Functions

机译：平衡批处理：语义分割损失函数的优化方法

获取原文
页面导航
  • 摘要
  • 著录项
  • 相似文献
  • 相关主题

摘要

Class-imbalanced data easily generates under-fitting problems in deep neural networks, which seriously limits the performance of the network. Several schemes have been proposed to alleviate the class-imbalance, i.e., data augmentation and network structure optimization. Our work has two main contributions: First, we proposed an optimized method Balance-batch which invests equal consideration for each class in mini-batch and tries to balance the losses of classes. Optimizing loss functions is a simpler and more effective way to tackle the class-imbalance than others. Second, we extend the binary classification Dice loss which is employed in medical images segmentation to multiple classifications. Multi-classifications dice loss reflects the improvement of Balance-batch to unstable loss functions. Our various loss function experiments on Pascal VOC2012 show the effectiveness of Balance-batch, which is beyond the advanced level of these loss function methods.
机译：类别不平衡数据很容易在深神经网络中产生拟合欠的问题，这严重限制了网络的性能。已经提出了几个方案来缓解类别不平衡，即数据增强和网络结构优化。我们的作品有两个主要贡献：一，我们提出了一种优化的方法平衡批次，这些批次投资于迷你批量中每班的同等考虑，并试图平衡课程的损失。优化损失函数是一种更简单和更有效的方法来解决比其他类的类别不平衡。其次，我们扩展了在医学图像分割中使用的二进制分类骰子损失到多个分类。多分类骰子损失反映了平衡批次的改善到不稳定的损耗功能。我们对Pascal VOC2012的各种损失功能实验表明了平衡批量的有效性，这超出了这些损失功能方法的先进水平。

著录项

相似文献

  • 外文文献
  • 中文文献
  • 专利
获取原文

客服邮箱：kefu@zhangqiaokeyan.com

京公网安备：11010802029741号 ICP备案号：京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有

  • 客服微信

  • 服务号