Deeplearning toolbox data input

I have hyperspectral data cubes of 640*480*100. And want to use patches of my data 16*16*100 as an input for the deeplearning toolbox, but i can't seem to get to get it working. The toolbox only allows image formats, but saving the patches as images isn't possible due to the spectral dimension.

[Cris LaPierre]

Could you be more specific? The Deep Learning Toolbox is a collection of functions, apps, and examples for designing, training, and analyzing deep learning networks. Which one of these are you having difficulty using?

[George navash]

Hi Cris,

I mean the deep network designer from the toolbox. It only allows jpg, png, etc as input format which consists out of mainly 3 dimensions, r,g,b. But my spectrum consists out of 100 dimensions for the wavelength aspect. I have saved my data in matrices of 16*16*100 but i cant use .mat as an input for the deep network designer

Answer 1

[Matt J]

but i cant use .mat as an input

Yes, you can.

imds = imageDatastore(location,"ReadFcn",@load);

You also don't necessarily need to keep your input data in files. The input to trainNetwork could be a 16x16x100xN numeric array, assuming you're using a recent enough version of Matlab.

[George navash]

Hi Matt,

I tried this and i get the following error.

I assume the @load is loading my .mat files as a structure. I am not sure how to solve this issue.

[Matt J]

Something like this, then,

imds = imageDatastore(location,"ReadFcn",@(z)load(z).yourImage);

[Matt J]

Incidentally also, you should probably look at randomPatchExtractionDatastore.

[George navash]

This worked thankyou!!

Answer 2

[Abhishek Tripathi]

We have one example for Classify Hyperspectral Images Using Deep Learning. Please go through this.

download our Image Processing Toolbox™ Hyperspectral Imaging Library MATLAB®

For more info. see the documentation here.