How to provide a score to an image based on pattern information in it?

Hello,

I am very new to image segmentation and pattern recognition. I have been reading quite a lot of literature and tried different functions available in MATLAB to find out patterns in the input image. I tried using Color-Based Segmentation Using K-Means Clustering, applied the Image segmentation tutorial and some other functionalities. These work well in segmenting the image and giving me details on the number of pixels, mean intensity of regions, etc. I was wondering if there is something like a numeric score which can be assigned to an image which contains more patterns in it (I tried using entropy as a score but it does not provide me a satisfactory measure to rank multiple images based on the amount of information in the image). Any information about already available functions/literature in the direction would be very helpful for me to proceed. Here are the images which I am using, and would like provide each image a numeric score based on the amount of information (patterns) they contain.

Answer 1

[Image Analyst]

There are a nearly infinite number of patterns in every image. Any grouping of pixels could be considered as its own pattern. You're going to have to be more specific as to what you would like the score to represent. How about the area fraction of pixels brighter than a certain threshold? How about the mean gray level of pixels brighter than a certain threshold? How about the standard deviation of pixels brighter than a certain threshold? We don't know what you want or what would be useful to you.

[Novice Geek]

Thank you for your prompt response. To start with, I have thousands of images which belong to a single dataset and each image is generated from a matrix of intensity values. Some images show prominent patterns (by pattern I mean a collection of high intensity pixels of some shape small or big) and some images just have random pixels distributed all over the slice area, without providing much information. Something like this:

My initial goal is to classify the images with some kind of pattern from these noisy images (like the above). Done that, I would like to see which images out of the rest have more prominent patterns, as to how a human eye would classify two images (as you said with pixels brighter than a certain threshold). From the 3 images which I have provided in my question, all the three have certain defined areas of high intensity pixels. For example, if I look at the following image at http://www.filedropper.com/mz69500 (sorry, I have used up my limit of file uploads here), I would give this new image a lower score as compared to the three images shown in my question, although a blue coloured pixel pattern can be clearly seen in the image.

Am I a little specific here?

I am not sure how would standard deviation of pixels brighter than a threshold help here. As shown in the images in the question, the blue regions, although they are of a lower intensity than the green regions of high intensity, so the two regions would have a different standard deviation value and how would this help in generating a score?

[Image Analyst]

Brighter pixels can be seen easier so both the mean and stddev would seem to correlate with human perception.

[Novice Geek]

Thank you for your response. I have a very basic question (due to my very limited knowledge). In your previous answer you have mentioned about a selecting brighter pixels above a certain threshold. How do I determine this threshold? I tried applying graythresh and multithresh functions to find a threshold level before calculating the standard deviation of pixels brighter than this threshold. But, the graythresh just provides me the green high intensity regions in the binary image. And, on looking at the code in your image segmentation tutorial, I see that the threshold has been set to 100 gray levels. I am very confused as to how should I determine a threshold which could be used for all the images in the dataset and then proceed to calculate the stddev for an image. Thank you very much for your help and time.

[Image Analyst]

You just have to decide how bright is bright enough to be considered. Maybe it's everything with a gray level of 0 and up, or maybe it's 1 and up. Or maybe 50 and up. I don't know how dark things get in your image and what should be ignored or paid attention to.