Thank you for providing nice toolbox. Just I have one question.
I am using pre-trained alexnet model to know the change according to signal-to-noise ratio.
My input image is rescaled to have small range (e.g., 108 to 148), and subtracted by "net.meta.normalization.averageImage". In this case, the accuracy falls down up to almost zero.
Visually, the input is well recognized, but CNN cannot.
Why this happened? CNN can not handle this kind of variance? or just problem of Alexnet?
Do I have to zscoring? or other normalization methods?
Please help me.
Maybe try to change the range just a bit? (e.g. [10 ... 245]) where the network should work fine. To see if the issue is in your code or in the network...
Hi,
Thank you for providing nice toolbox. Just I have one question. I am using pre-trained alexnet model to know the change according to signal-to-noise ratio. My input image is rescaled to have small range (e.g., 108 to 148), and subtracted by "net.meta.normalization.averageImage". In this case, the accuracy falls down up to almost zero. Visually, the input is well recognized, but CNN cannot.
Why this happened? CNN can not handle this kind of variance? or just problem of Alexnet? Do I have to zscoring? or other normalization methods? Please help me.