Hi,
I want to find correlated regions present in an image to implement copy-move/image splicing detection in an image for forensic application. Presently I have divided an image into overlapping blocks of size 8x8 and extracted the SVD features from each overlapping block and placed them in a single row and it become 1*992016 dimension for one image. Similarly, I have extracted such features for 10 images and formed a feature matrix of dimension 10*((256-8+1)*(256-8+1)*16) to train the data for SVM. But it's taking a lot of time to train the SVM and I am obtaining inaccurate results. So can you suggest for my following questions?
1. is there any techniques to further reduce the dimensions of the feature matrix without losing its significant information.
2. Can I train the SVM classifier with a feature matrix of one image after another image(individually).
Dear Prabhu Bevinamarad ,
Look the link, maybe useful.
https://www.mathworks.com/matlabcentral/answers/518589-using-singular-value-decomposition-for-feature-extraction-from-images
Regards,
Shafagat
Thank you for the suggestion... I went throught the reference...In my implementation, I have already extracted features of image using SVD. I was further thinking about reduction in feature. The suggested reference doesn't serve my purpose but gave little idea about other technique how to incorporate. Any way thanks for suggestion mam.
And Can I train the SVM classifier with a feature matrix of one image after another image(individually).?
As we know, the use of SVD introduces three matrices 𝑈, 𝛴 and 𝑉. But, not all the elements in these matrices have the same degree of importance. The non-zero elements of the diagonal matrix 𝛴 are arranged in descending order (𝛴11≥𝛴22≥.....≥𝛴𝑘𝑘), which represent the singular values of the input matrix.
After obtaining the SVD of each block, you can take only the top left element of 𝑈 matrix (𝑈11) and the first two non-zero elements of 𝛴 matrix (𝛴11 𝑎𝑛𝑑 𝛴22) to represent the block features. Thus, each block can be implemented by three numbers (elements) only. According to the obtained results, you can add more elements. As a result, the size of features' vector is highly reduced in case of using the SVD.
For more details, please refer to the following article:
Article Face Recognition using Maximum Variance and SVD of Order Sta...
Thank you for your suggestion sir.. and I will implement the code and verify...
Why not make the svm programmable to take the image, then you would not have to re-invent the wheel making double job, in a format that is compatible to be converted to other formats using Google converters, remember essentially all data is only Ones and Zeros, they all speak the same language.
Thank you for your suggest sir. And currently I am using SVM for training and testing image features. But I'm not getting good results. So is there way to improve the results by changing the SVM model parameters? I mean is there any parameters present in SVM model which play significant role in classification accuracy.
I have 256x256 forged input image and my objective is to find forgery region present in an image to find correlated pixels present in an image. Therefore I am dividing input image into overlapping block of size 8x8. Thus, I am getting total 62001 blocks. And for each overlapping block I am applying SVD and taking diagonal elements. like wise I am obtaining SVD values for each overlapping block and placing column wise so for 1 image I am getting 62001x8 values 4,96,008 values. Like wise I am extracting features for 20 images. Where each row represents an image feature so in total my feature dimension is 20x4,96,008. So, when I train this I am getting following difficulties.
1. It is taking lot of time for training.
2. Giving in accurate results.
3. The size feature dimension has become very large.
So, Is there any other technique
1. to reduce feature dimension.
2. To change SVM model parameters to improve accuracy.
3. Any other technique to replace entire model.
I am feeling very happy in this research community because I am getting lot of suggestions, recommendations, and new ideas.
Thanks for your support
Thanks for your valuable suggestion mam. You mean for each overlapping block PCA is applied and only one heighest Eigen value should be selected?
- Badges
- Science topic
- Similar topics
- Computer Science
- Images