I have been using Euclidean distance as the method for calculating the difference between any two images. However, my search for better result using other distances (such as Mahalanobis and Cosine) yielded the same results. which makes me wonder;
1. what are the advantages of one over the others?
2. and are there other similarity calculations that are more suitable for comparing vectors representing images?
Hi Dear
I hope the attached paper sufficient your quastion
thanks
Try Pearson Correlation Coefficient, which is one of the best measure.
See for example it use and normalization in
http://www.springer.com/cda/content/document/cda_downloaddocument/9781447124573-c2.pdf
Hi Dear
I hope the attached paper sufficient your quastion
thanks
It depends on the application. The question is therefore Why do you need to compare the images? Is it related with quality assessment or maybe is it related to image similarity and retrieval?
You can find two examples of this in the links below as well as similar question in research gate.
I hope this is useful. Further details can be useful if you want to get more detailed answers from the other researchers in the community.
http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=1453520&queryText=euclidean+distance+images&newsearch=true&searchField=Search_All
http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=5710601&queryText=Objective+Video+Quality+Assessment+Methods%3A+A+Classification%2C+Review%2C+and+Performance+Comparison&newsearch=true&searchField=Search_All
https://www.researchgate.net/post/How_can_we_measure_similarities_between_two_images
Similarity or dissimilarity? Same thing really. For one you find the maximum, the other you find the minimum.
My understanding is that the Pearson Correlation Coefficient and the Normalized correlation coefficient are the same thing. They are nice because they confine the metric to the -1 ... +1 interval, which makes setting thresholds a bit easier. Also, these 'normalized' methods disregard signal power, so well-matched weak signals score just as well as well-matched strong signals. Most of the time that will be what you want, but not always.
I posted a similar question recently but didn't get much of a response. As people have already pointed out there are so many methods to do this. I was looking for other methods that didn't fit into my simple (direct or transform) taxonomy.
For image processing, if speed is important, I reckon sum-of-absolute differences would win easily. It too can be normalized to give a metric between 0 and 1. But a frequency-domain approach might be better if you are looking for correlation in a specific frequency band - so you can essentially do filtering at the same time as correlating.
Similarity measures can be compared in terms of simplicity, speed, dimnesionality, immunity to outliers and noise, etc.
I found a new metric which is invariant to data scale, noise and outliers, simple and fast.
It is perfect for CBIR systems, a student of mine is making excellent results using this distance, you may find the paper at:
https://www.researchgate.net/publication/264995324_Dimensionality_Invariant_Similarity_Measure
Article Dimensionality Invariant Similarity Measure
I too found a new similarity measure. I'm sure there are many more to come!
Conference Paper A New Statistical Measure of Signal Similarity
Interesting work Hugh.
I need to study it thoroughly... and compare to my work. perhaps we can do something together.
have you prove it as a metric? and if it is invariant to signal shifting, such as the DTW?
One of the points of difference about the method I describe is its ability to handle an arbitrary number (M) of channels.
I did that work more than 10 years ago. I would do it a bit differently now.
Dear Abass,
Euclidean distance (ED) is a simple summation of pixel-wise intensity differences of images. It is straightforward, easy to code and fast to calculate. However, it is blind to correlation, sensitive to outliers, and small deformation may yield a large ED which is not supposed to be. Mahalanobis distance can solve partially theses issues since it considers the covariance in calculations.
For robust and accurate results you may try to use kernel function. Here you can find more details about it:
http://crsouza.blogspot.com/2010/03/kernel-functions-for-machine-learning.html
The answer lies in the calculation of SSIM (structural similarity index).
Read the paper by Wang "mean squared erro" love it or leave it!!
- Badges
- Science topic