about calculating a measure of an image feature vector

Hello,

I want to ask a question about calculating a measurement of an image feature vector. For an image, several feature were extracted and the feature vectors in different length were obtained. For example, I have a feature vector of 256 dimension. However, for some reason, I have to calculate one measure to represent this feature vector. Could anyone here give me any advice? Thanks.

ZG.

Answer 1

[Image Analyst]

You need to have some other metric to compare your feature vector to. For example let's say you want to estimate how attractive a person's face is, and you have these features: distance between eyes, eye color, hair color, skin color, number of blemishes, width of mouth, size of nose, hair style, distance between mouth and center of eyes, sex, and age. You can't just give one number for all those features, but you can build a model to give one number (estimated attractiveness) based on a weighted sum of those. But to derive the model you need something else, like human visual scores for the images. Then you can build the model to try to predict the human score.

[ZhG]

Actually, these features are only texture features of the image. I extracted each kind of them independently. So I obtained a set of feature vectors in different length. But I need to calculate one measure to represent a feature vector. For example, standard deviation of one feature vector or entropy of it. But I am not sure that whether my idea is correct. I mean that I don't know whether it is correct to calculate a measure in this way.

[Image Analyst]

You can't. They each tell you something unique and useful. Can you describe what a person looks like with just one number? Well, sort of, but not as accurately as if you use all of the feature measurements. Why do you want to get down to a single number anyway? What good does it do you? All it does is throw away useful information. Standard deviation of the numbers in your feature vector is a totally and completely useless number and you should not use it.