I am working on palm-print recognition project, and I want to ask how I can extract the ROI accurately.
in the following example (image 1 ROI) I could extract the ROI from the image using image opining using SE=50*50,
however, in some other images, a part of the arm appeared and couldn't be removed by the same SE which make problems in the feature extraction phase.
Any suggestions to exclude the arm or to extract ROI accurately?
Thank you in advance.
1-Find the middle finger
2- find its regression line
3- expand the line towards the hand by the length of the finger or + something , perhaps 125% or so of the length of the middle finger.
4- P= the perpendicular line by that end
5- remove all pixels located at the right of P
6- optimize parameters that suite your database
another solution:
a) if the hand was horizontal:
1- A = a segmented flipped version of the hand
2- B= a segmented flipped version of the hand
3- C=A && B
4- use the right most white pixel of C to cut the original hand from the arm
b) otherwise:
1- find the direction of the hand (regression line of the whole hand can be appropriate
2- P=the perpendicular line at the end of the middle finger
3- flip the hand around P
4- do the same as in a)
I suppose you are able to extract the hand (with arm attached, if present) from the rest.
You could then do a distance-transform: each pixelvalue will be the distance (Euclidean or Manhattan) to the background. Then find the pixel will the highest distance, and crop the foreground with a circle of certain diameter (just as much a 'guess' as the 50x50 kernel you used for the closing operation) and that found pixel at its center.
Maybe a kind of spherical horse solution, but you might get away with it.
Dear Lambert, if the sizes of the hands are different, and this is quite possible, a certain diameter of a circle, will not work well. Same applies for the morphology operations.
Hi Ahmad, glad to meet here again.
True, the sizes of the hands are different. You could use the highest Euclidean distance value as the radius of the circle. Maybe add or subtract a fraction of it. I think it can work and the method certainly is easy to implement.
imregionalmax(), max(), find(), bwboundaries(), hypot(), etc.
For hand gesture image analysis, see the published literature on it for algorithms:
http://iris.usc.edu/Vision-Notes/bibliography/contentspeople.html#Face%20Recognition,%20Detection,%20Tracking,%20Gesture%20Recognition,%20Fingerprints,%20Biometrics There is a list of links to papers:
21.4 Gesture and Pose Analysis
21.4.1 Human Posture, or Human Pose, Human Body Pose 21.4.1.1 Human Pose Tracking, Posture Tracking, Posture from Video 21.4.1.2 Human Posture, Upper Body Posture, Arms 21.4.1.3 Animation and Reconstruction of Human Model from Posture, or Pose 21.4.2 Gesture Recognition Techniques 21.4.2.1 Gesture, Overviews, Surveys, Evaluations 21.4.2.2 Pointing Gestures, Pointing Systems 21.4.2.3 Hand Gesture Systems, Hand Posture, Hand Pose 21.4.2.4 Hand Tracking for Gestures 21.4.2.4.1 Arm Tracking, Arm Pose for Gestures 21.4.2.5 Gesture Recognition Systems, Applications, Human Computer Interaction Systems 21.4.2.5.1 Gesture Systems, Surfaces, Desks, Tables 21.4.2.6 Music Related Gestures, Systems, Music Video Analysis 21.4.2.7 Sign Language, ASL Recognition 21.4.2.7.1 Sign Language, Fingerspelling
Actually, I chose another way, I tried to fit the maximum square inside the binary and the max squre will be in the palm area (see the Sample attached 23,24 and 25) this method will work in good manner without caring about the arm and any other unwanted regions .
but the problem is that this method is not rotation-invariant which mean if there any rotation in the hand this will shift the square a little bit to the wrong area (see example (30))
is there any method to to standardize the rotation of the hand (with the arm) before applying square fitting
suggestion please.
Quote Ahmed: " is there any method to to standardize the rotation of the hand ".
That's easy: find the two intersections of the ROI with the image-border!
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing