-1

according to the below link, the writer has implemented the image segmentation with vgg16 as the encoder and the random forest as the classifier. https://github.com/bnsreenu/python_for_microscopists/blob/master/159b_VGG16_imagenet_weights_RF_for_semantic.py

As I know we need to take care of two concepts when implementing U-net : 1- Upsampling the extracted features in the decoder part 2- Adding residual connections to preserve the actual values of the input image.

Question 1:

How we can use only random forest instead of the decoder part of Unet?

Question 2:

I follow the tutorial video corresponding to the code(on youtube). The teacher sent the image to the encoder part and reshaped the extracted feature from (8,1024,996,64) to (8153292,64).

I am confused by this statement:

He says we have 64 features.

I think we should have 8153292*64 features. because I think any pixel is a feature

ali ali
  • 1
  • 1

1 Answers1

0

Q1. VGG16 is used to encode the image (contracting path of the U-Net). Once the features extracted, you can use any classifier you want to classify the pixel as 0 or 1, or you can reconstruct a segmentation mask using the expansive path of the U-Net. The strategy used in your link is to use a random forest classifier.

Q2. After the contracting path of the U-Net, each of the 8153292 pixels is now represented by 64 features, instead of 3 (R,G,B).

Klinsmann
  • 33
  • 4