SIFT features extraction from images and send it to CNN or other classifiers

Good day,

I have been strugling for days with the following code:

1. imds = imageDatastore('mydata','IncludeSubfolders',true,'LabelSource','foldernames');
2. numImages = numel(imds.Files);
3. for i = 1:numImages
4. img = readimage(imds,i); img = im2gray(img);
5. img = imbinarize(img); % points = detectSIFTFeatures(img);
6. hogFeatures(i, :) = extractHOGFeatures(img,'CellSize',cellSize);
7. end

........... The above code is part of MATLAB help, and works well for my data that contains arabic alphabets one per image. (sample attached).. I have little code to convert these features to an .csv file from a code of SVM classifier or a CNN reads features from the .csv file and does the classification.

When I want to do the same for SIFT features, instead of HOG, the code becomes as:

----------------

1. imds = imageDatastore('mydata','IncludeSubfolders',true,'LabelSource','foldernames');
2. numImages = numel(imds.Files);
3. for i = 1:20
4. img = readimage(imds,i); img = im2gray(img);
5. img = imbinarize(img); points = detectSIFTFeatures(img);
6. SiftFeatures (i, :) = extractFeatures(img,locations);
7. end

.................................line 6 generates and error that this type of indexing is not allowed in this type of objects.I can see the SiftFeatures object has complex nested list of variables (7 varibales ) of different dimensions and their inner dimension vary with every images.

my questions are:

1. How to convert this data in a feature vector to save it as .csv file or any other suitable form a CNN can read it.
2. What features I must take (scale, orentation, octave ..... etc) for better classififcation, or I need to pick all of them.

thanks in advance.

Shahid B.