Title: U-Net Convolutional Networks for Biomedical Image Segmentation
Authors: Olaf Ronneberger, Philipp Fischer, Thomas Brox
Link: https://arxiv.org/abs/1505.04597
Quick Summary:
One of the interesting things about the paper is the architecture used. They share they weights of the “contracting path” in the “expansive path”, and since the sizes are different, they are forced to crop them (blue dots in the layers).
One of the challenges faced by the authors were the separation between different cells that belong to the same class. For this, they propose using a weighted loss where those separations have a large weight in the loss function.
Training uses a high momentum (0.99) because they don’t have many training samples. Thus, a large number of the previously seen training samples determine the update of the current optimization step. In addition, they perform data augmentation for the model to be rotation and shift invariant, and robust to deformations and gray value variations.
