3 ms·
>> According to Google, "[Image Search] analyzes the text on the page adjacent to the image, the image caption and dozens of other factors to determine the imag
by YeGoblynQueenne 10y ago
>> According to Google, "[Image Search] analyzes the text on the page adjacent to the image, the image caption and dozens of other factors to determine the image content."
The reference for this (missing from the op) seems to be a paper titled "Evolving data mining into solutions for insights" [1]. I don't have time to read this now and the one-liner abstract is not very informative.
Still, it dates from 2002, whereas allegedly Google Image Search nowadays uses actual image processing. There was a different article submitted on HN earlier today that basically says Google's machine learning algorithms used in Image Search now "understand" what the image content is [2] (which is patently b.ll.cks but, hey).
My best guess is that Image Search does use more modern machine learning but still operates largely in the old model, where it looks primarily in the text accompanying an image- and in fact even treats that text as a good, old-fashioned "continuous bag of words" rather than more modern techniques like word embeddings, or even N-grams.
Edit: basically what I'm saying is that Google seems to take in a query like "unprofessional hair" and return all documents that have the words "unprofessional" and "hair" in them, even when "unprofessional" is qualified with "not" very clearly (the retrieved documents are tiny).
That seems a bit weird- even Google web search seems to be better than that at taking some context into account.
[1] http://dl.acm.org/citation.cfm?doid=545151.545174 http://dl.acm.org/citation.cfm?doid=545151.545174
[2] https://news.ycombinator.com/item?id=11975878 https://news.ycombinator.com/item?id=11975878