How Image Quality Affects Deep Neural Networks

Recently I read a paper from Samuel Dodge and Lina Karam Arizona State University. It is talking about how the noise in image may considerably affect DNN's classification result. And usually these kind of distortion won't seriously affect human's judgement.

In most machine learning competition, people only feed in good quality images, DNN achieved very impressive result compare with those human manually extract feature machine learning algorithms. But when we add some noise into these images (in reality blur or low quality images are very common), the success rate of cross validation test would decrease significantly.

So it would be a very interesting research topic to find a way to make the current DNN network more stable and e invariant to these distortions.

Here is the link of the paper