Hello,

The number of clusters is the visual words index, you have to calculate the number of occurrences for each index to get the histogram (features vector).

For the normalization do you mean the feature scaling.

I hope this help you

https://en.wikipedia.org/wiki/Feature_scaling

a useful link is given blow where you may find the method to construct a histograms of user provided  data-sets.

https://gist.github.com/alfredplpl/6901429

I think this tutorial can answer your question.

http://people.csail.mit.edu/fergus/iccv2005/bagwords.html

when you get 500x128 dimension matrix, this means that the size of your word vocabulary is 500. Each 128 dimenstion of the matrix is a vector of word. For a image which can be represented as Xx128, bow simply count the numbers of each word. That is first compare the each descriptor from X descriptors with the 500x128 matrix, and get the index of matrix which is the most similar(least euclidian distance), and count the number of the index of all X descriptors.

As you mentioned, you have 500 visual words (here each word corresponds to a line of X), and thus the histogram is about the word occurrences of these 500 words. That is a statistical process. For a given image, you can surely get a certain number of subimage patches by some sampling strategy (dense sampling, etc.). For each patch, you can extract a 128 dimenstional feature vector. Then, by similarity measurement, the feature vector of this patch can be mapped to one of the 500 words by minimum similarity distance (e.g.Euclidean distance). For all the patches in an image, repeat the above, and you will get the corresponding words for all patches. By histogram statistics, you can get the histogram. 

For normalization of histogram H = [h1,h2,...h500],just do the following to make the sum =1.

H_norm = H/sum(H)

Thank you for all, at last i got the answer it's working well, special thanks to Lijun Zhao. also can some one tell me recent research trend in " funds image retrieval using CBIR'.