nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

Find Tool

bradstew

Thanks for the help with my previous question.

I am attempting to use the "Find..." functionality as a tool for identifying products. The idea would be to take photo with a digital camera of the unidentified product, run it through a predefined color correction task, and run the find tool against a catalog in media bin. I am wondering if anyone has attempted to do something similar to this before, and what kind of success people have had.

I find it to be very picky in regards to the size of photos. It would return products that look nothing like the photo if it was somewhat similar in pixel size.

The one successful test that I ran required me to edit the image in Photoshop by making the image identical in resolution and pixel size and performing some adjustments for color correction. I found though, when I attempted to turn these changes into a task in Media Bin (with the transformation tools), it produced a similar (though higher contrasted) image which was unsuccessful in turning up a correct image.

Any suggestions/comments/thoughts?

Find more posts tagged with

Comments

Migrateduser

My take is that the Content-Based Image Retrieval (CBIR) algorithm used by MediaBin is heavily influenced by color percentages.
I was told that it's a complete black box (even to them), but my guess is that the image is broken into several blocks and color histograms are generated, then compiled into some kind of signature.
Signatures are then compared and a similarity value is returned.

So color is very important, and emplacement of the color "spots" comes in second (try to match an image against a color-corrected version of itself and a rotated version of it and you will see what I mean).
Saying that color is very important, means that the background is also important... If the background is very large compared to the pictured object, the algorithm may be considering two images as similar because their backgrounds are.

If you can do a screen copy of one of your typical search results, I may be more specific on why I think some images are being retrieved.

Obviously, I would be more than interested in hearing any official word about the used algorithm

Folco Banfi

bradstew

After doing some more experimenting this morning, I came to basically the same conclusion as you. The whitespace was contributing to some of the poor results. By integrating some size conversion and color correction tasks, I came up with a standard that appears to work quite well with my sample product. It'll require some more testing, but I've been very satisfied with the results.

What would be interesting to see is the matched percentage (like in search engines). Is it actually recognizing specific colors, patterns, and locations for my product at a high 90+% rate, or is it returning the result simply because it has a similar amount of whitespace?

lyman

The papers which formed the basis for the Interwoven (formerly Iterated Systems) Content-Based Image Recognition (CBIR) system were published at an SPIE conference. Here are the references. The abstract and full text in PDF are available from http://spie.org/app/publications/ (unfortunately for a fee but technically the journal and not us "owns" the copyright).

Similarity-based retrieval of images using color histograms
Keshi Chen, Stephen G. Demko, Ruifeng Xie

Publication: Proc. SPIE Vol. 3656, p. 643-652, Storage and Retrieval for Image and Video Databases VII; Minerva M. Yeung, Boon-Lock Yeo, Charles A. Bouman; Eds.
Publication
Date: Dec 1998

Image descriptors based on fractal transform analysis
Stephen G. Demko, Mehdi Khosravi, Keshi Chen

Publication: Proc. SPIE Vol. 3656, p. 379-389, Storage and Retrieval for Image and Video Databases VII; Minerva M. Yeung, Boon-Lock Yeo, Charles A. Bouman; Eds.
Publication
Date: Dec 1998

See also:

United States Patent 6,697,532 Demko et al. System and method for producing descriptors for image compression

Despite the name it really concerns content-based search.

The simplest conclusion is, yes it uses both spatial and colorimetric descriptors.

Cheers,

Lyman