The noise pattern used in testing is fixed, because randn('seed',0); is always called before adding the noise: https://github.com/cswin/WIN/blob/master/Demo_test_WIN5RB.m#L51
Further, the input image is always resized to 321x481 inside Consistency.m. This means that the exact same noise pattern is used in every test. The accuracy drops significantly if I comment out randn('seed',0), or if I resize the input image to something other than 321x481. This suggests the provided CNN models are overfit to encode this specific noise pattern.
The noise pattern used in testing is fixed, because randn('seed',0); is always called before adding the noise: https://github.com/cswin/WIN/blob/master/Demo_test_WIN5RB.m#L51
Further, the input image is always resized to 321x481 inside Consistency.m. This means that the exact same noise pattern is used in every test. The accuracy drops significantly if I comment out randn('seed',0), or if I resize the input image to something other than 321x481. This suggests the provided CNN models are overfit to encode this specific noise pattern.