For image processing developments I'm interested in a descriptor (or rather: a routine) that characterizes the structuring of a displayed object in a sensible manner, if possible by a smaller number of single values - e.g. some numbers describing maxima in spatial frequency occurrences or the like. The numbers describing the image structure should not be sensitive to the objects orientation in the image.
For clarification: if you have an ordered set of matches (to be considered as a combined object) displayed in the image, the desired descriptor should show the same values regardless of the orientation of the set of matches in the image. But in case the matches are evenly and chaotically spread across the scene the descriptor should yield different value(s), as the structure of the image has changed.
An answer would be most helpful if based on/related to the available routines of the open source image processing library ImageMagick. But any answer is greatly appreciated.
Hello Roosen, Please go through herewith the attached files, i think these may helpful to you.
Please see if following on google is of help .
A GENTLE INTRODUCTION TO THE BASIC CONCEPTS ...
noodle.med.yale.edu/hdtag/notes/shape_notes.pdf
by HD TAGARE - Related articles
Cheers
@Suresh Merugu: Thanks a lot for sharing your files! I'll give them a thorough reading. But just at the first glance the pointer to the SIFT feature extraction procedure immediately opened up a vast area of additional information, even for a combined search on "imagemagick" and "SIFT". This certainly will help me to proceed!
Apart from that, any further pointers are gladly appreciated nonetheless! :-)
Central rotational moments or more generally moment tensor. If you scale these by zero-order moment, you'd get moments that are also scale invariant.
Thank you very much so far for the contentful responses to my question. After a quick scan of the propsed papers and methods the "bag of words" approach seems to be most interesting for me, but I'll reflect on the other literature as well.
Just to clarify a bit more what I am searching: I'm looking for an approach to classify images of objects that belong to certain classes that need to be distinguished. I'm not oriented towards a near-to-identical search. As an example (probably better than my introducing one with the matches) take the following one: Think about cars and lorries and motor bikes, photographed in a mostly similar perspective and with a quite distinguishable background. A basic goal would be to classify the images w/r/to the gross vehicle classes (bikes, cars, lorries, etc.). A certainly more ambitious goal is to classify for something like production generations of same passenger car brands that are discernible by slight variations of silhoutte, headlamp layout, or front hood design and the like.
In my real application (not cars) color is at least a partial discrimination objective. Therefore I already hit upon the PHASH functionality of my preferred software base ImageMagick. The first results are promising, but the PHASH value is not addressing any structural features, imho. Therefore I am researching additional trait retrieval approaches.
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing