Hello everyone,
I captured some images from of hyperspectal imaging system. My final aim to identify the material composition from the images. However before that i need to perform some image preprocessing like
calibration, noise removal, feature extraction , dimensionality reduction. can you please suggest some code or routine to the do these task?
Also after this task, i wanted to apply classification algorithm so it can detect the material from it
Can you guys tell me a way to do in the Matlab. Thanks a ton.
Kristen Chappel there are many examples of codes for preprocessing your data available on the internet. However, if you are planning to work in ML and DL I suggest you go for python instead of MATLAB as the ML community in python is larger than MATLAB which can be a great help if you stuck somewhere.
hi kristen,
first off, just don't do it in matlab. I did similar processing for my phd work in matlab. I eventually got the results I needed but it was very very slow. also we did not have a lot of alternatives back then. I am aware things got faster and better within the last 10-15 years but matlab should not be your choice for processing large multi-dimensional data. have a look at python libraries. if you have coding skills definitely go with c/c++. most of the stuff you will be doing is linear algebra and there are very good (fast and accurate, gpu accelerated) libraries - free.
regarding your hyperspectal images, are they reflectance or radiance images? if they are radiance images, do you have an accurate model of your illuminating source?
next processing. you mention
1. calibration: see above. regardless of the sw tools you prefer, you will also need a good model of your sensor.
2. noise removal: I don't have much experience in this but you should probably check IEEE transactions papers for "hyperspectral image denoising". also I know it sounds cheesy but google is your friend.
3. feature extraction: see above. also, I would definitely check CNN based approaches. I have been seeing many papers on this recently.
4. dimensionality reduction: ICA and PCA. PCA is pretty much trivial unless you pick a really bad sw platform.
5. classification: your should probably start with Gaussian mixture modeling and use some thresholding based on an appropriate distance measure (mahalanobis maybe?) for classification. if you believe you have discriminative features, SVM should be your starting point. if you do not want to go down the classical way, there is some active research on applying deep learning on hyperspectral data (restoration, super-resolution, ...) you should definitely read.
good luck.
- Badges
- Science topic
- Similar topics
- Papers