Note_Automatic Water-Body Segmentation From High-Resolution Satellite Images via Deep Networks

Basic Information

Section 2 article,
Automatic Water-Body Segmentation From High-Resolution Satellite Images via Deep Networks

notes

Point of departure

1. Water segmentation is the basic task of remote sensing.
2. Traditional methods rely on spectroscopy and can only handle low-resolution images. High-resolution images contain more detail.
3. The robustness of the method is tested by obtaining data from different data sensors.

main innovations

1. Propose a new segmentation network RRF DeconvNet network (restricted receptive field deconvolution network)
2. It is considered that the usual loss, that is, the L2 distance or the Euler distance, cannot well highlight the role of the boundary in the segmentation. Therefore, EWLoss is used, which is the Gaussian weighted Euler distance.

Detailed description

Network structure information and network configuration.

It can be seen that the main changes are in the network structure, and a large number of hole convolutions are used to replace ordinary convolutions to obtain a larger receptive field.

Using Gaussian weighted distance as the cost function, the main characteristic idea is to highlight the edge, and the closer to the segmentation boundary, the greater the weight obtained.

experiment

The data is extracted from google earth, 0.5M, the size is 512*512, mainly in Sichuan, Wuhan, a total of 9,000 pictures, 7:2:1=train:validation:test

Mainly six experiments, (two types of loss functions, three networks)

There are also two evaluation indicators, one is overhead pixels and the other is edge pixels. The former is conventional

\[OP= \frac{TP+TN}{TP+TN+FP+FN}\]

The latter one is not very clear, the approximate description is this (pixels with a maximum distance of 5 checkerboards from the border are considered edge pixels):

 the pixels who have a maximum 5 chessboard distance to boundaries are considered as edge pixels

Looking at the results, it seems that the evaluation function has little effect on the results, or basically has no effect. OP is the mainstream evaluation index. Why is the improvement not obvious? Obviously, because there are too many pixels in the picture and too few pixels in the border area, there is a bit of sample imbalance, so even if it is improved, there will be no obvious changes. . Therefore, the introduction of new evaluation variables, the purpose of evaluating loss is achieved, and it can indeed make the edge more accurate and clear.

Summarize

1. Change the network structure
2. change loss
3. Design favorable evaluation indicators
4. It is worth mentioning that the framework used is mxnet