The notes of paper U-Net: Convolutional Networks for Biomedical Image Segmentation.
Big Question: classification in pixel level and thus image segmentation
reason
- single label for a image is not enough to support segmentation. ### background
- why focus on it
- biomedical images: like cells segmentation. Add: also appropriate for other entity segmentation
- how have been done:
- the development of deeper CNNs
- using CNNs segmenting electron microscopy images
- what have been missed:
- computation efficiency and redundancy: slow cause every patch require a running of network; patch overlapping
- difficult trade-off for localization and the usage of context.
Methods
For what?
Cells segmentation.
Framework of Methods
Convolution2D + deconvolution (upsampling 2D). The output of one downsampling layer is contracted as part of the input of the corresponding symmetric upsampling layer.
Novelty
- data augmentation randomly elastic deformations: shift, rotation, gray value, random elastic deformations are the most important
- replace pooling by upsampling.
- No fully connection layers.
- weighted the loss of touching objects (cells).
Details
weighted map to segment overlapped cells
According to the paper, they pre-compute the weight map for each ground truth segmentation to compensate the different frequency of pixels.
data augmentation
Smooth deformations using random displacements vectors on a coarse 3 by 3 grid. The displacements are sampled from a Gaussian distribution with 10 pixels standard deviation. Per-pixel displacements are then computed using bicubic interpolation.
Abstract
The main idea in abstract are contracted NNs and data augmentation so that the new NNs can get reasonable results by fewer images.