Hello,
I am working on Ultrasound images and calculating gray level co-occurence matrix (GLCM) and gray level difference matrix (GLDM) at four orientations(0,45,90,135 degree). Here, suppose we find Contrast at 0 ,45,90,135. Now, is there any specific reason why we need to average the Contrast of all the directions and take it as only one feature or we can take it as four different features. And is the averging of feaure has to be done in both the cases i.e. GLDM and GLCM. Please guide.
I think you must used each of four feature separately because by average you destroyed separability of feature so I suggested to concatenate four feature
Averaging in this case was done perhaps to make the feature invariant to rotation.
In some applications, invariant features are important to generate stable results despite having transformation to original source images. Otherwise, averaging is not mandatory.
Best
Hello,
The reason is to prevent a directional-bias. However it is not mandatory, I personally have used both forms.
Regards.
Read my paper. Here I explain when you can average the result for different angles and the correction to be made if you want to average.
https://www.researchgate.net/publication/312953835_Improvement_of_co-occurrence_matrix_calculation_and_collagen_fibers_orientation_estimation
Conference Paper Improvement of co-occurrence matrix calculation and collagen...
In many applications, one could be interested in a rotation invariant descriptors and/or scale invariant ones. This is the reason why some people prefer averaging of different angles. So if the orientation is important in you application, you should neglect these things.
In addition, one can try other than simple averaging. For example, non-linear operations such as max/min/mid often show better performance. One can also try other statistical features beyond average, such as variance and higher order moments. Hence, it is really up to the problem how the "features pooling" is designed
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Human-Computer Interaction