Skip to content

This project focuses on dividing the IBSI dataset using the UNet network and delves into the two modules of up-sampling, while using a very intuitive visualization tool to be able to view the results more easily. The code has been annotated, and was written on December 8, 2023, and recently uploaded to GitHub because it was needed.

Notifications You must be signed in to change notification settings

lovekang3344/IBSI-Segmentation

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

UNet对细胞壁的分割(Pytorch)

项目主要采用Pytorch实现,并使用tensorboard作为直接可视化工具,方便看到整个模型的训练过程。 架构.png

该项目主要使用UNet网络对ISBI数据集进行分割训练,并对不同的模块采用对比的方式比较效果。 主要对比的两个模块是

  • ConvTranspose
  • Upsampling

以下测评数据均在 $epoch = 6$ 的结果获得。

模块\指标 loss mIoU mPA result(若无法跳转,请直接下拉至底部)
ConvTranspose2d 0.3512 0.8832 0.9438 效果1
Upsampling 0.3404 0.8874 0.9640 效果2

这里从结果上看,由于IBSI数据集过于简单,Upsampling表现更佳,在既没有增加额外参数的同时 推理时间性能 均优于ConvTranspose2d

但是 转置卷积图像噪点去除的性能上优于 线性上采样Upsampling(这一点我们可以对比分析效果1和效果2的动图)

效果查看

终端中输入如下命令,并在浏览器中访问 http://localhost:6006/

tensorboard --logdir='./logs'

Loss

这里我使用的 Loss 都是 CrossEntropyLoss loss_compare.png

mIoU

mIoU_compare.png

mPA

img.png

ConvTranspose2d

ConvTranspose2d_28.gif

Upsampling

Upsampling_28.gif

About

This project focuses on dividing the IBSI dataset using the UNet network and delves into the two modules of up-sampling, while using a very intuitive visualization tool to be able to view the results more easily. The code has been annotated, and was written on December 8, 2023, and recently uploaded to GitHub because it was needed.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages