Can we use Unet to differentiate on pixel level?

Can we use Unet to find whether each pixel is belongs to which class? One class is a matrix containing all '8''s and other other containing all '30's stored in separate MAT files.The label image for class 1 is a white image equal in size to the matrix of class 1 stored in Mat file and label for class 2 is a black image. The semanticseg results in traning, validation and testing accuracy of 100%. I saved the trained network. When I then give input of all '8' or all '30's matrix to this trained network, and check the probabilities on 'Softmax' layer by using activation function of the saved trained network, the probabilies are extremely low for class#1 and very high for class#2, no matter what the matrix contain, whether all 8's or all 30's.

Please help me. I shall be very thankful to you.