PageRank for Images
https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=4522561
This paper discusses how to apply Google’s popular PageRank to image search. In particular, it extends Google’s PageRank algorithm and uses a simple assumption: images found within a website are relevant to the topic. In other words, if many images contain the same object, then it can be said that there is a theme common to all these images. Thus, the paper’s extended PageRank algorithm for images involves finding themes in a given dataset of images and seeing how common they are throughout.
For background, in class, we discussed he mechanics of the original PageRank algorithm. Essentially, a website is more “popular” if there are many other websites linking to that one site. Furthermore, if a “popular” website will carry more weight to its “endorsement” if it is linking to other sites.
This is different from traditional PageRank in which the original algorithm primarily deals with text-based search. In traditional PageRank, images can be found but many of the images from the search results are irrelevant. The article describes how using the extended PageRank algorithm for images have improved the search results. More specifically, the article describes how many irrelevant images within the search results were removed once the extended PageRank algorithm was used. Interestingly enough, while the original PageRank algorithm relies on using the number of links to generate top search results (as discussed in lecture), this extended PageRank algorithm for images does not utilize links. Instead, this extended PageRank algorithm has utilizes a “visual link” – which is like the link idea used in original PageRank – which is essentially the probability that an image will be visited by someone.