Consulting on Tensorflow project for image segmentation

I need consulting on a Tensorflow neural network I have written. I have a neural network written in Tensorflow for image segmentation (a simple binary classification -- foreground or background). It is three convolutional layers connected to 3 deconvolutional layers. I can get the program to converge in training and the test images predict with 0.99 accuracy. However, predictions are complete noise. The current state of the project is outlined in this SO pos: [login to view URL] I just need some consulting as to why predictions are complete noise? I have attached an image of a prediction. I have also attached the program file -- [login to view URL] is the network and [login to view URL] is the data file. the [login to view URL] file goes in the directory [login to view URL]