the following paper
P. Vincent, H. Larochelle Y. Bengio and P.A. Manzagol, Extracting and Composing Robust Features with Denoising Autoencoders, Proceedings of the Twenty-fifth International Conference on Machine Learning (ICML‘08), pages 1096 - 1103, ACM, 2008.
contains a variety of images on the MNIST dataset that show how well features are recognized when different levels of noise are added. Especially important are the last pictures where he shows that the more noise is added the better the network learns dependencies between variables. With low noise levels features do not stand out.
The link to the paper is the following: http://www.iro.umontreal.ca/~lisa/publications2/index.php/attachments/single/176
Also note that the noise that is added is not really white noise or anything similar. It is simply setting 20,30,50% of the values at random to zero.