How to do feature extraction from an image?

Hi I want to do feature extraction from an image. I read a paper and did this steps: I did image segmentation. Then I want to do feature extraction. In this paper:

Segmented lungs were divided into 3*3 windows in which all nine pixels were located in the lung mask. Window size selection is a compromise between higher resolution (in the classification process) and faster algorithm. Smaller windows (i.e. 1*1 or 2*2) have the problem of more time complexity for training and increaseing the number of FP. Larger windows (i.e. 5* 5 or larger) cause lower resolution of reconstructed image after classification and miss some tiny nodules. Thus, for better resolution and faster algorithm, simultaneously, we used a 3*3 window. In the training process, these windows were labeled as nodule (þ1) and non-nodule (1).

My question is this: Is there any standard criteria to lable the 3*3 window as a noudle? ( I mean if how many of these pixcles are 1, we should lable the window as a noudle?)</pre>

[Rik]

Questions like this are very specific to a field. To my knowledge (and judging from your description) there isn't a standard. Often people experiment to find what works best. I suspect the authors of the paper you're citing (but not naming, bad habit to get into) simply tried a few window sizes and found that 3x3 worked well enough without requiring too much computational power.

Have a read here and here. It will help you judge if a question is relevant on this forum, and if so, how to best write it.

[sara]

Thank Rik. So I should try multiple tries to find best method of this problem:)

[sara]

[sara]

I mean a process like top image.

[Rik]

If each block is independent of the others, you could use blockproc to do the processing. Then you only need to decide how to reduce each block to a single value.

[sara]

Yeh. My problem is this: how to reduce each block to a single value

[Rik]

That is a domain-specific question. I don't know what makes sense in your situation. Some usual methods are max, min, and mean (with or without a threshold).

[sara]

Thank dear Rik. I tried them.

[Image Analyst]

What did they do to segment the image before the filtering?

And after they filtered the segmented image with a 3x3 window, or any size with any values, what did they do with the filtered, segmented image? What values were in the filter kernel window?

[sara]

Hi They segment image from background and then they labeled 3*3 windows with 0 and 1. I tried to do this with this picture that I attached. Thanks

[Image Analyst]

I don't think they segmented the image. I think they did that on the original gray scale image. I don't think it would make any sense to do a covariance of 9 pixels if the 9 pixels were segmented, which means they are already binary/logical.

[sara]

thanks dear Image Analyst. I think I made a mistake. I read the paper again. I knew they segment region of interested from background and then they do this operation on gray scale image.