Image enlargement degrades performance

Question

I am working on an image segmentation project and am trying to apply an image enlargement technique to increase the size of a set of tutorials. At first, I just try to use horizontal flip to double the size of the image, but I found that performance is much worse than not using it. Is there an understanding that can be used. Thank.

Answer 1

So basically you need to answer one important question: Is the inverted image a valid image in your domain?

• If not, then it can harm your learning process simply because you provide the network with invalid input that can examine your network false patterns in your data. It is not so rare that flips can harm your learning - for example, when recognizing a logo, it is important not to change the orientation of your data in order to study the logos correctly.

• , , . , , . -, - , , . , , , , .

  : , .