I have a classifier for two classes that was trained in a convolutional neural network using cuda-convnet. Input are hand-labeled greyscale images, it is part of a scene-labeling project.
What I'm trying to achieve with the classifier is for it to find (what it considers to be) errors in the hand-made labels of a dataset, i.e. it should classify every pixel in a greyscale image, and if its estimation differs from the label of that pixel, that area (and only that area) should be marked.
How can this be done? I don't now of any algorithm/method/scientific paper on the subject. Any suggestions are hihgly appreciated.