Adding Gaussian noise to train images. Is this the right way ?

[Erogol]

#################################
## Add Noise
#################################


layer{
  name: "gauss_noise"
  type: "DummyData"
  dummy_data_param{
    data_filler{
      type:"gaussian"
      std: 5
    }
    shape:{
      dim: 64
      dim: 3
      dim: 32
      dim: 32
    }
  }
  top: "gauss_noise"
  include {
    phase: TRAIN
  }
}


layer{
  name: "noise_drop"
  type: "Dropout"
  bottom: "gauss_noise"
  top: "gauss_noise"
  include {
    phase: TRAIN
  }
  dropout_param {
     dropout_ratio: 0.9
  }
}


layer {
  name: "add_noise"
  type: "Eltwise"
  bottom: "data"
  bottom: "gauss_noise"
  top: "data"
  include {
    phase: TRAIN
  }
  eltwise_param {
    operation:SUM
  }
}

I tried the above configuration to add gaussian noise to image but my net start to behave strangely after which I suspect that this is not the way. I tried to create gaussian noise with std 5 and drop 0.9 of them and add the noise to initial images. Do you think something is missing ?

[li kai]

I think this is right, waitting for someone to answer

在 2015年9月8日星期二 UTC+8下午11:36:15，Erogol写道：

[Bharat Bhusan Sau]

Hi, 

What problem did u face ? Is it solved ? Please let me know as I am also going to add noise in my dataset in this way.

Bharat Bhusan