Denoising-autoencoder on validation set give bad result and even for train image randomly noised at instance

Question

I trained a small denoising autoencoder ConvNet on STL-10 dataset and the output on train/test _noisy data is clean. However if I take a validation image and add noise and pass through the network it still has some noise in the output.

Also one KEY observation is when i take any train set image and add noise to it on run time and pass through the network it also has poor performance output as validation set image. The pictures of output are attached.

Any advice is welcome!

Answer 1

Be sure you have a healthy split of data for your validation set; it sounds like it's becoming too rigid. As for architecture, consider using MaxPool2D layers to pull key features, and then restore them later down way with UpSampling2D. Hope that helps!