In my research , I want to do image pre-processing. So I would like to know which is better for the face and ear image preprocessing.
Hard to say; whatever filter works best; maybe blurring also works or maybe you don't need a smoothing filter at all.
A Gaussian preserves more features like edges. You need to play around with kernel-sizes as well. Just try!
(my gut feeling is : unsharp mask, or a Gaussian followed by a Laplacian)
What is your objective behind the pre-processing? Please tell us more on what you want from your filter.
Both the filters you have mentioned are low pass filters, so looks like you want some kind of smoothing effect to remove noise.
Median is used in case of salt-&-pepper noise whereas Guassian filter is like taking weighted average and hence removes signal components in high frequency.
Both the filters will blur the images and so like Dr. Lambert suggested you will need some sharpening.
Why are you not considering Eigenfaces? Let us know your application so that we can give precise pointers.
Regards
As Ernesto Moya-Albor well said above, the advantage of median is that it helps getting rid of outliers.
those kinds of filters are application dependent, you will never know unless you have some prior info about your images.
as Lambert Zijp said "Hard to say; whatever filter works best" "
Therefore, I suggest to do adaptive median/mean filtering, check to see if the area is homogeneous, i.e. almost similar intensities do mean else do median.
In my experience the Gaussian Filters are the best choice in pre-processing image stage but I work in a controlled environment. I think you need to review which are the characteristics of the environment in which the image is going to be taken and use the best filter for those conditions
Gaussian Filter is better choice, have look on it please
www.ccis2k.org/iajit/PDF/vol.9,no.6/3778-9.pdf
Both the Gaussian filter (equivalently a weighted arithmetic mean filter) and the median filter are only appropriate under "ideal" noise circumstances, which never occur in real images. We can always construct such ideal circumstances within the lab to test our algorithms, but ultimately determining an appropriate approach in a given application will often require domain specific knowledge.
However, if we want to consider the practical question of which operator is more robust in general, under various noise conditions, then you should consider that there are a large number of alternative filters based on computing local (region-based) averages and many of these will likely serve better in real applications.
As a starting point for investigating robust filters, you might consider the following paper, which directly addresses the evaluation of several monotonic and non-monotonic averaging functions (including the mean and median, as well as several robust estimators) in the context of pre-processing of noisy images in a facial recognition task. This paper also showed that scale reduction by averaging is also a very effective denoising technique that, when using robust averaging operators, not only eliminates degradation due to noise, but can actually improve the performance of your recognition algorithm as compared to "clean", non-reduced images (by clean, I mean "no additional noise added, only that already present due to the imaging system").
http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=6622458&tag=1
I don't know if you are still looking at this, but I have recently written a paper on a type of smoothing/de-noising filter that I am particularly fond of ...
Conference Paper Improved IIR Low-Pass Smoothers and Differentiators with Tunable Delay
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing