Kaggle-2018-data-bowl

Record of the competition

1.Data Pre-Processing

Histogram Equalization

It makes the image constrast more obvious. It maybe helpful in preprocessing.

Cropping Strategy

This strategy is aimmed at seamless segmentation of arbitrary large images. Missing input data is extrapolated by mirroring.

2.Deep Learning Models

U-net

U-Net is the basic Network in this competition. It has played an import part in many projects. Here, the number of filters is reduced for fasting training.

M-net

M-net is proposed in Jan,2018 by Huazhu Fu. It adds some structure on the base of U-net and tries to solve the information loss in pooling and upsampling.

3.Results Display

Here are some of the segmentation results. The red line draws the segmentation results of the colorful images after gray scaling.