Image segmentation semantic accuracy of measurement Methods
There are three ways to measure the accuracy of the main image semantic segmentation:
- Pixel accuracy (pixel accuracy, PA)
- The average pixel accuracy (mean pixel accuracy, MPA)
- 平均 IOU (Mean Intersection over Union, MIOU)
Before introducing the three methods, we need to explain the meaning of some notation.
: Total number of categories, including the background if the word is
: Real pixel category of pixels is predicted to be the category of the total number, in other words, for the category of pixels, the error is divided into categories the number of how many.
: Real pixel category of pixels is predicted to be the category of the total number, in other words, for a real category of pixels, the total number of pixels divided on how much.
1.PA
PA's significance is very simple, and our conventional classification accuracy rate calculation makes no difference, that is, the total number of pixels divided by the total number of pixels.
2.MPA
MPA is an improvement to the PA, which is calculated for each class to PA, and then averaged for all classes PA.
3.MIoU
Semantic segmentation, MIoU is the standard method of measurement accuracy. It is separately calculated for each class (AC prediction result and real tags and ratio) IOU, IOU and then averaging for all categories.