I made a deep network which its input is one image with 224224 pixels.The network output is a one image with 224224 pixels.
for Correction of weights in training stage i calculated intensity difference between the pixel by pixel in output image and refrence image, with MSE.
I reached loss=0.0019 and acc=0.3682 after training stage.
I need the help for improve of accuracy.

My expectation from the network is that the pixel-to-pixel intensities of the network output image and the reference image are close to each other. But my level of accuracy is apparently very low.
In the first step, I want you to tell me the errors in my code.
In the second step, I don’t know what is the criterion for calculating accuracy in this network and how can I change this criterion and customize it.
i used:
activation function is “relu”(in the middle layers)
activation function is “sigmiod”(in the end layers)
padding is same
optimaizer function is “adam(lr=0.001)”
loss function is “huber”