Solved – Image normalization for object detection

Question

I am training object detector using CNN.

I have read lot of articles, that images should be normalized using (img-img.mean)/img.standardDeviation

However, is this really neccesserry? Using this the images become really confusing and its hard to see objects i am trying do detect.

Also is it recommended to just normalize data and traing net on these datas, or add normal data to this normalized data?

Answer 1

Neural networks with inputs centered & scaled to have mean 0 and variance 1 tend to be easier to optimize. This is because the scaling amounts to preconditioning the network.

It's not surprising that displaying the centered & scaled images is hard to interpret -- these images are not encoded in the usual way images are for human consumption. But that's fine! Your neural network doesn't need to care about that -- it just needs a consistent representation that it can understand.

Also is it recommended to just normalize data and traing net on these datas, or add normal data to this normalized data?

I'm not sure what "normal data" you would add to the images. It's standard to train the network on the centered & scaled images.