Image segmentation - improved network structure

Image segmentation - Improve the network structure

1. How to find questions worth asking ？

• Low model accuracy
• How low is the accuracy of the model
• The accuracy of a certain class of samples is low

Near the edge of the picture mask High deletion rate ; Specific shapes such as rectangles mask Poor shape prediction ; The class with less uneven samples in the training set has low accuracy

• There are structural problems in the forecast ：mask Low recall rate ; The segmentation accuracy is improved , The classification accuracy is reduced

Example 1：

• The recall rate of positive samples is very low （ see notebook、model_visulize.ipynb）
• Multi task learning （Muti-Task）
• Modify the code 、 Running experiments

Example 2：
problem ： Whether there is a better network structure ？
Way 1： Increase network volume , turn up neuron Number .going deeper
Way 2： Better network architecture ：Vanilla Unet, SCSE Unet,Attention Unet

2. Various tasks

• Classification task ：encorder+classifier layer（1 layer ）
• Split task ：encorder+decoder

encoder/backbone:
resnet family ：resnet series ,resnext series ,se-resnet series ,se-resnext series
efficientnet family ：b0,b1,b2,……,b7
other ：inception series ,vgg series ,densenet series

efficientnet series

• More efficient （ The network size is relatively small ）
• Higher accuracy

3. Multi-stage Training program

By freezing - Partial network （freeze）, Play a fine adjustment ：

• The migration study
• Small amount of business data
• It's noisy

Multiple picture sizes ：

• First in 128*128 Training on small pictures
• Finally, by increasing the image size finetuning

Use different loss Conduct finetuning

• First use BCE Training
• Last use lovasz loss/Dice loss, Fine tune a few EPOCH