** First of all, in your case, there are a whopping "26" classes!!! Which means that you need a multi-class SVM (handles more than 2 classes).

** "svmtrain" implements a binary SVM (P.S: svmtrain is not available in the newer versions of MATLAB - instead uses 'fitcsvm'). So, you need to use "fitcecoc" which implements a multi-class SVM. Please refer the MATLAB doc link: https://in.mathworks.com/help/stats/fitcecoc.html.

** Secondly, the error is because the SVM function expects the training data (in your case "trainingset") to be row vector format i.e. in your case, there are 486 features and 2600 observations. Here, 1x486 (not 486x1 !!) is a row vector where the columns represent different features. So, the final trainingset matrix will be of size: 2600x486.

** Now, for each observation, there should be a class label. This is necessary to build a model using SVM (Training: where you tell the machine- "this is how my classes are going to behave!!! ").

The class labels must be given in the variable 'tar' and it is a column vector. In your case, the size of 'tar' must be 2600x1, where, each value is a label for corresponding observation (row vector).

Good luck!!

** First of all, in your case, there are a whopping "26" classes!!! Which means that you need a multi-class SVM (handles more than 2 classes).

** "svmtrain" implements a binary SVM (P.S: svmtrain is not available in the newer versions of MATLAB - instead uses 'fitcsvm'). So, you need to use "fitcecoc" which implements a multi-class SVM. Please refer the MATLAB doc link: https://in.mathworks.com/help/stats/fitcecoc.html.

** Secondly, the error is because the SVM function expects the training data (in your case "trainingset") to be row vector format i.e. in your case, there are 486 features and 2600 observations. Here, 1x486 (not 486x1 !!) is a row vector where the columns represent different features. So, the final trainingset matrix will be of size: 2600x486.

** Now, for each observation, there should be a class label. This is necessary to build a model using SVM (Training: where you tell the machine- "this is how my classes are going to behave!!! ").

The class labels must be given in the variable 'tar' and it is a column vector. In your case, the size of 'tar' must be 2600x1, where, each value is a label for corresponding observation (row vector).

Good luck!!

Thank you so much, my problem has been solved. Could you help me more, I have used Mdl = fitcecoc(B,A) where B is training dataset(2600x486) and A is target value 2600x1. what should i do next to check accuracy. and how can I plot confusion matrix.

I am familiar with ANN there "net = train(net,trainingset,tar);" used for training and "Result = sim(net,trainingset)" used for result.

Please, I'll be grateful if you can help me further.

Thank you

Well, what you have done now is - built an SVM model for the given training data with labels.

**Now, you need to have a test data (different from those given for training) for this purpose.

**On this data you would test the model you have trained.

** For this, you need to use the function "predict"

eg: predict(mdl,test_data);

For more details refer the following link: https://in.mathworks.com/help/stats/compactclassificationdiscriminant.predict.html

** Confusion matrix and accuracy can be computed as follows:

** confusion_matrix = confusionmat(testing_labels, predicted_labels);

For more details refer:

https://in.mathworks.com/help/stats/confusionmat.html

** accuracy = sum(testing_labels == predicted_labels) / numel(testing_labels);

Hi Abhilash,

Thank you so much.

My problem has been solved.

Hi

Abhilash K Pai I am doing the same work as Neha. I am doing the image classification. after doing fitecoc I hane a model "Mdl". I want the labels to show my labeld image. how can I do this?

thanks in advance

Samreen Mughal Sorry. I didn't understand your question!

Abhilash K Pai the problem is solved thanks for your response but I have one more question

my image is 100*100 and i have 27 features for every pixel. Total number of pixels is 10000. final matrix is 10000*27. I want to classify image with 3 classes with name 1,2,3.

I don't know what will be the class label? I know the size of label matrix should be 10000*1. should it be like {1,1,1,1,.......,2,2,2,2,......3,3,3,3,3}?

In this way I got very bad results. can you help me?

If X is feature vector and Y is class label.

In am coping some lines of my Matlab program

t = templateSVM('Standardize',1);

Mdl = fitcecoc(X,Y,'Learners',t,...

'ClassNames',{'1','2','3'});

my question is about "Y" in program.

thanks in advance.

Samreen Mughal

a) For "each" image: 1) Features are extracted and 2) "A class label" is assigned (either 1 or 2 or 3 in your case).

b) Then the function "fitcecoc()" is used to create an SVM model (stored in "Mdl"). This is known as Training.

c) Finally, you use the model to test how it performs on different data other than the ones which is given for training (using function "predict") This is known as testing.

*** The key here is that the labels (Y) need to be known for training the system.

The labels depend on what you are trying to do and hence if labels are not available you can manually assign a corresponding label to an image according to your application.

Abhilash K Pai thanks for your response.

I have one gray scale image for segmentation into 3 classes. for example I have 3-by-3 pixel image. total 9 pixels. every pixel has 10 features. Now feature matrix is 9*10.

class label Y should be 9*1. Now the correct labels for each pixel are as:

{2 3 2 1 3 2 1 1 2}.

I am not gating that should the class label must be correct label? or I place them as:

Y={1 1 1 2 2 2 3 3 3}?

If I use class label as Y={1 1 1 2 2 2 3 3 3}, it gives some results but not good enough (I got 50% error rate).

if I use class label as Y={2 3 2 1 3 2 1 1 2}, it gives meaningless results.

What should do?

Thanx in advance

Samreen Mughal : According to what I can understand, you have to segment a given image into 3 classes. Am I right?

Abhilash K Pai yes. I am just confused about class labels.

Samreen Mughal : You should train your system based on features.

** i.e, According to your example each pixel has 10 features.

** based on the values of this feature you will have 3 possible classes.

** hence these values need to be trained along with their labels (pixel-by-pixel), not the entire image.

** So, while testing also you need to consider pixel by pixel and try the model (you trained) to predict the label for each pixel.

Now, this becomes a totally complex process.

Abhilash K Pai thankyou so much for your kind response. I will try this.