Presenting Compressed Information

Greg Hughes posted a very interesting video on his blog. It is a presentation from SIGGRAPH on novel ways to re-size an image. Normally, an image is simply scaled: every nth pixel is removed, and this uniformly alters the image size. It can also be cropped, which results in complete loss of information contained in rectangular regions of the image.

The SIGGRAPH presentation proposes a third way. The importance of certain pixel paths (from top to bottom or left to right) is determined. The least important paths eliminated as and when the image is shrunk.


Many years ago, I invented a technology that used the same principle on text (US Patent #6,172,685). To present text in a smaller space, it assigned importance to words and eliminated less important words. The visual impact was not quite as dramatic, but it was interesting to see the “gist” of the text preserved and communicated in far fewer words.

Avidan and Shamir had to find some way to measure the importance of pixel paths. Similarly, I had to find some way to measure the importance of words. At the time, I was immersed in the study of information retrieval, which provided solutions like tf-idf. However, such solutions completely ignored the meaning of the words. My technology used text that had been parsed into grammatical structures. The subject (word or phrase) and verb (word or phrase) of each sentence could be treated as most important. Adjectives and adjective phrases and prepositions and prepositional phrases could be treated as less important. The less important words could then be eliminated, and a useful summary of the text could be presented in fewer words.

As far as technology to successfully and accurately parse the text into grammatical structures, I left that to some one else. I demonstrated the technology with text that had been manually parsed.

Avidan and Shamir addressed a similar problem: pixel paths that went through faces were more important than any automated algorithm would reveal. They showed a tool to manually circumscribe faces and assign paramount importance to those pixels. They also mentioned automatic face recognition algorithms could be used.

Tags: Language, Technical

Updated at: 30 August 2007 12:08 AM