How can I improve my NN results?

I selected the training values from corrupted parts' neighbors, and use them as a training data(input data). Target data is their neighbors' pixels.

TRT(2*4574): 162 163 164

37 37 37 ...

kkT(1*4574): 0.341176470588235 0.341176470588235 0.325490196078431 0.356862745098039 ...

net=newfit(TRT,kkT,200)

net=train(net,TRT,kkT)

output=net(TRTIT)

I take some results:

>> output=net(TRTIT) %prediction

output =

0.4295

>> I(158,40)%original

ans =

0.4235

How can I improve my results? Is that training good for image inpainting case? I am really looking forward your answers, Regards,