What does Predict for neural network give back?

I'm built a unet for segmentation purpose. I trained the network and now want test the accuracy of the network. For training I randomly extracted 32x32 patches with randomPatchExtractionDatastore. For testing I split the input image into ordered 32x32, so 516x516 images was split into 256 32x32 images. Then I run the following code

testdata = imageDatastore(testDir);  
predictPatchSize = [32 32];
net = load('trained_unet.mat');
net = net.net;
YPred = predict(net,testdata);

And the result it give back to me is a 4D array of 32x32x2x516.

Now I get the 32x32 because that's the size of my images. the 516 is how many images I put in and now many I get out. However the 2 is confusing me and I can't extract out the 516 images to rebuild back into a 516x516 image to compare to the grand truth mask I have to see how accurate the network is. Anyone have any idea on this problem?

Answer 1

[Jon Cherrie]

I think that Ypred(i,j,k,n) is the score for class k for pixel i,j of the n-th image.

Often the score is the probability or similar. So you can think of this as the probability the pixel i,j of the n-th image is in class k.

[Zehao Pan]

Hmm but I'm trying to segment the image. Like it should just be giving me back just the mask of the area that suppose to be segmented?

Hmmm oh I see my mask has 2 class like either the background or the thing I'm trying ot find. So reason it has 2 probability is because there are 2 things it could be. So I would need to compare them to see which one is more likely and just set it myself?