diff --git a/docs/imgs/dice.png b/docs/imgs/dice.png new file mode 100644 index 0000000000000000000000000000000000000000..56f443dfade0a02240dad61d6554a23c91213bb5 Binary files /dev/null and b/docs/imgs/dice.png differ diff --git a/docs/imgs/dice1.png b/docs/imgs/dice1.png deleted file mode 100644 index f8520802296cc264849fae4a8442792cf56cb20a..0000000000000000000000000000000000000000 Binary files a/docs/imgs/dice1.png and /dev/null differ diff --git a/docs/imgs/dice2.png b/docs/imgs/dice2.png new file mode 100644 index 0000000000000000000000000000000000000000..37c3da1f1906421c0d3928ab18212a4d1a0966a0 Binary files /dev/null and b/docs/imgs/dice2.png differ diff --git a/docs/imgs/dice3.png b/docs/imgs/dice3.png new file mode 100644 index 0000000000000000000000000000000000000000..50b422385ee1e6b0cf7652ac63571652ce1d52ef Binary files /dev/null and b/docs/imgs/dice3.png differ diff --git a/docs/loss_select.md b/docs/loss_select.md index 454085c9c22a5c3308c77c93c961628b53157042..f7e975822c50efdcc1e80cadb40fbc5f55fafb1e 100644 --- a/docs/loss_select.md +++ b/docs/loss_select.md @@ -11,24 +11,35 @@ 我们从DeepGlobe比赛的Road Extraction的训练集中随机抽取了800张图片作为训练集,200张图片作为验证集, 制作了一个小型的道路提取数据集[MiniDeepGlobeRoadExtraction](https://paddleseg.bj.bcebos.com/dataset/MiniDeepGlobeRoadExtraction.zip) -## softmax loss与dice loss - -在图像分割中,softmax loss(sotfmax with cross entroy loss)同等的对待每一像素,因此当背景占据绝大部分的情况下, -网络将偏向于背景的学习,使网络对目标的提取能力变差。`dice loss(dice coefficient loss)`通过计算预测与标注之间的重叠部分计算损失函数,避免了类别不均衡带来的影响,能够取得更好的效果。 -在实际应用中`dice loss`往往与`bce loss(binary cross entroy loss)`结合使用,提高模型训练的稳定性。 - +## dice loss dice loss的定义如下: -![equation](http://latex.codecogs.com/gif.latex?dice\\_loss=1-\frac{2|Y\bigcap{P}|}{|Y|+|P|}) +
+
+
-
+