I have read some recent articles about unsupervised recognition of objects. So, I began to think about unsupervised learning and tried the following algorithm for an unsupervised bag of words applied to images: 1) Divide a set of images into small patches 2) "Vectorize" the patches and build a matrix X with the vectors 3) Find the principal components (PC) of X, and project X over PC 4) For each patch build a label with a binary alphabet (1, 0), one letter for each PC where 1 if the projection of the vector over the respective PC is great or equal to 0, and 0 otherwise. 5) Apply NLP methods to the doc built of labels assigned to a given image, in the same sequence of the respective patches.
I got some results: the verification? of the Zipf law, the recovery of the most similar image based on tf-idtf, identification of "stop words", and others things I have to review and understand
Do you think this is an interesting research theme?
[link][6 comments]