I am writing a program in Matlab using CNN concept and want to create output classes based on ground truth data i have collected. Input to Neural network is array of Image....How to create output for using BPN for classification?
Dear Parminder Kaur:
In order to compare, register and fuse images, you need to describe them in a normalized way. Hence, you will need to apply some transformations to them before starting your data mining stage. The MATLAB website has lots of information on this.
There is a remote sensing toolbox. Please, refer to the following links:
https://github.com/IPL-UV/altoolbox
https://www.mathworks.com/matlabcentral/fileexchange/41768-validation-toolbox--performance-metrics-for-remote-sensing-data-and-climate-model-simulations-?requestedDomain=www.mathworks.com
http://www.mdpi.com/2072-4292/6/7/5938
http://www.mdpi.com/2072-4292/6/9/8617
Regards,
V. V. Estrela
Dear Parminder Kaur:
In order to compare, register and fuse images, you need to describe them in a normalized way. Hence, you will need to apply some transformations to them before starting your data mining stage. The MATLAB website has lots of information on this.
There is a remote sensing toolbox. Please, refer to the following links:
https://github.com/IPL-UV/altoolbox
https://www.mathworks.com/matlabcentral/fileexchange/41768-validation-toolbox--performance-metrics-for-remote-sensing-data-and-climate-model-simulations-?requestedDomain=www.mathworks.com
http://www.mdpi.com/2072-4292/6/7/5938
http://www.mdpi.com/2072-4292/6/9/8617
Regards,
V. V. Estrela
Thanks @Vania.
But my question is after this normalisation process. How to create target classes using ground truth/ROI data for classification for output layer of NN?
Firstly you are feeding images to NN or features of images.
if you are using features of images so then you have to give label to each image during feature extraction phase to each image in separate array with corresponding location of image number.
if your feeding images directly then you can count number of each images in each class and make separate array of ones, twos and threes as much the number of classes
In your question, it was not clear at which point you were in terms of processing. Assuming you describe the features using the same type of feature vector as Mr. Amin Ullah said, you have to calculate the distance between the feature vectors according to a similarity metric or metrics.
The distance between the feature vector of the ground truth and a perfect match will tend to zero.
Regards,
V. V. Estrela
Thank you @ Amin and @ Vania..
I am trying to do per pixel classification i.e. why my feature is DN values/ reflectance values.
@ Vania "The distance between the feature vector of the ground truth and a perfect match will tend to zero." is formula to find output class..
Use "save" for workspace variables . https://www.mathworks.com/help/matlab/ref/save.html
Generally , as Vania has mentioned in numerical (R2) space , the distance metric is commonly used and they are pretty static (as in do not change ) .
Commonly pdist give the similarity and dissimilarity of the vector data ...
Dear Parminder Kaur:
For each pixel, you have several physical attributes. Let us say you have the following features for each pixel located at (m,n):
I(m, n) = Gray level
R(m, n) = Reflectance
T(m,n)=Type of soil
F(m,n) = Modified Normalized Difference Water Index (MNDWI)
A possibility for a feature vector fv could be fv(m,n)=[I, R, T, F], where each entry is a number. Now you can choose the metric. Seek in the literature one that is convenient for your problem. Examples: Manhattan distance, Euclidean distance, correlation, L1, etc.
https://en.wikipedia.org/wiki/Feature_selection
http://web.mit.edu/mobility/publications/istvs1.pdf
Regards,
V. V. Estrela
- Badges
- Science topic