DeepLearning: Inspired by DSS ——Multi-layer merge loss and Relu

最新推荐文章于 2024-05-28 09:39:45 发布

原创最新推荐文章于 2024-05-28 09:39:45 发布 · 238 阅读

0 ·

CC 4.0 BY-SA版权

hao's DeepLearning blog 同时被 2 个专栏收录

6 篇文章

订阅专栏

DeepLearning

3 篇文章

订阅专栏

本文探讨了Deeply Supervised Salient Object Detection with Short Connections论文中提出的深度监督方法。该方法通过整合多级得分层来计算损失值，加速了网络收敛，并在实验中表现出了更好的效果。文中详细介绍了网络结构，包括从卷积层到得分层再到损失层的流程，以及最终得分层的拼接。对比使用和不使用ReLU的实验结果表明，不使用ReLU的版本收敛更快。

Contact Me:
王雪豪 xuehaowang@buaa.edu.cn

In Paper <Deeply Supervised Salient Object Detectionwith Short Connections>, the author merge multi-level score layers to calculate the Loss value.

It used the structure shown as follow:

conv1 -> score1 -> loss1
conv2 -> score2 -> loss2
conv3 -> score3 -> loss3
conv4 -> score4 -> loss4
conv5 -> score5 -> loss5
conv6 -> score6 -> loss6

concat(score1, score2, score3, score4, score5, score6) -> score -> loss
INSTEAD OF
concat(score1, score2, score3, score4, score5, score6) -> Relu -> score -> loss

Meanwhile, I test the network with Relu and without Relu, it turned out to be better without Relu and converge faster.