U-Net

Image Segmentation

Input images

• Class label is assigned to to each pixel
• Input images are split to multiple patches so that it could have much larger number of input images than original datasets.

• Input of the U-Net has larger size than outputs. Because the input images has padding.

• The reason why the output class is two is because the datasets have two different classes.

Contracting Path

• Learning what is in the image?

Expanding Path

• Learning where the objects are?

Training

• Cross Entropy Loss maske stee output have two channels.

• Some people say that output could have one channel and will have sigmoid function to classify.

Prediction

• Overlap-tile is used when prediction.