Neural Network for snowscreen detection

I'm trying to use deep neural networks to determine if a 160x100x3 image is a snowscreened (just black or white randomized pixels) or not. I adapted the digit classifier example to handle color images and generated the training set by generating a 160x100x3 snowscreen or taking a random section of a phone's homescreen. The network diagram is attached, it is 2 autoencoders and 1 softmax layer. I've been trying for a couple weeks to get it to detect snowscreening but it really isn't all that accurate. Does anyone have any advice for structuring the network or specifying the autoencoder training options? I'm really not sure where to go from here except to keep changing size/specs of the autoencoders hoping it works.

Thanks!

[Greg Heath]

Downsize the inputs until you have figured it out. Sorry I can't help otherwise.

Greg

Answer 1

[SAM]

have u trained it successfully , i have the same problem it gives v good accuracy on training while performing worst than ever on test set