Neural Network Segmentation confused about the type of input and training.

Hi

I am trying to use the neural network toolbox for aiding in image segmentation. After reading alot of answers on how I can do this, I have done as follows.

My images are of size 640x480. I reduced their size to half [320x240] and extracted five different features on them namely, mean, std. dev, kurtosis, entropy and skewness. Each of these area images. So I use (:) operator to turn in to vectors and stack them beside each other. So if I had 5 features for one image then my input matrix is = [ 5, (320x240)]. Is this correct ?

Also for the target I have my ground truth, on which I have extracted the centers and the edges. And again used the (:) and stacked the three of them together as target matrix as such = [3, (320x240)]. Is this also correct ?

I use them to train with the nprtool and from there I am very confused how to get the output and use it to test other images ? I would appreciate any guidance in this matter please.

Answer 1

[Walter Roberson]

Your features are probably not going to be (320x240) in length. If the total length of the features is L, then you would end up with an (number of images) by L, or L by (number of images) array. It is not required that each feature type have the same length as the other feature types, only that each feature type have the same length for each image. For example, you could take standard deviation on each 160 x 120 quarter, creating 4 standard deviations, and those 4 together could be one "feature", by you might choose (for whatever reason) to take the mean of the image as a whole, so that might add just 1 (subtotal so far: 5), and so on. You might end up with 5 features with a total length of (say) 12, so you would have 12 by number of images as your input.

[Lubna Shahid]

Hi Walter

Thank you for replying me back. Is it enough to subdivide the image in 4 quarters and find features on them ? I have attached my image so you can have an idea of what I am working with. The other samples are similar to this. I am trying to get the blobs segmented.

[Walter Roberson]

Dividing in quarters would not be enough to train on. I am not sure what would be good to train on considering there are a varying number of blobs.

[Lubna Shahid]

Thank you for your reply again. I am uncertain as well, and that is why I thought of reducing the image size and just taking the whole of it to train. But it does not seem to be working correctly.

[Greg Heath]

When you present a matrix as input to a NN,

EACH COLUMN is interpreted as a single input corresponding to the same numbered column in the target matrix.

Therefore, input and corresponding target matrices must have the same number of columns.

Hope this helps.

Greg

[Lubna Shahid]

Hi Greg

Thanks for the reply. Can you please advise what kind of features would be best to extract from the image that I have in order to train the network to only detect the blobs in the image ?

[Greg Heath]

Sorry, image feature extraction is not my bailiwick.

Good Luck,

Grreg

[Lubna Shahid]

Thank you for the help Greg