Googles new Image Search function

[pvg photography]

<p>I just discovered Googles new image search function, where-by you can drag any JPEG image into the search box and do a search. While dropping in some of my own photos, I was surprised to find a couple of them on sites other than photo.net. What a useful tool for checking on copyright infringement. Anyone else with similar experiences?</p>

[MichaelChang]

<p>I've found many of mine use elsewhere too, mostly in Russian and Chinese blogs and photo sites. </p>

[funcrunch]

<p>Great tip - I just found one of my (copyrighted) images used on two Chinese blogs. Trying to decide whether it's worth trying to get compensation.</p>

[tom_mann1]

<p>Unfortunately, I am disappointed in the results produced by Google's new image search capability. It seems that their index of similarity is mostly color similarity. If you want a really good laugh, just drag my image of two beetles mating (attached) into their search box.<br>

;-)</p>

<p>Tom M</p>

<div></div>

[MichaelChang]

<p>Tom, at least it's all very consistent - Chinese Blog sites where these things are delicacies! LOL</p>

[tom_mann1]

<p>Michael - Someday, I'll tell you the story of what I found in the men's room on the night train from Chengdu to Guanyuan City in Sichuan. :-0</p>

<p>Tom M</p>

[pjmeade]

<p>I tried the image search on this image, which is one of my most "used" photos: - <a href="

I've been aware of some of the cases, but identified some new "users" and have sent off half a dozen emails to ask for it to be removed.<br>

It could be a good adjunct to tineye</p>

[randrew1]

<p>I'm not a professional photographer and my on-line presence is not huge, but I did find some of my pictures in use in Russian and Bulgarian travel blogs. I also found two of my images used by a local taxi company on a page listing local attractions. These images faithfully reproduced the watermark that lists my website. On one level I appreciate the advertising, but I still sent them an email asking that they contact me to arrange for permission to use the images. </p>

<p>I agree with Tom Mann that when the search doesn't find a near match, it finds something with similar colors. I found some examples as incongruous as Tom's, but not as humerous.<br>

I also found that the search engine choked on larger images (10 MP jpegs). I don't know what the size limit is. I don't know whether the practical limit is measured by pixels or bytes. </p>

[john_cogburn]

<p>I am curious about the new function, but I don't understand how to use it. I tried dragging some images into the google image search box, and the only thing that happened was that the browser opened to the html of that one particular photo. For example, I tried dragging your beetles into the google image search box (from one browser window to another), and the browser went to the html of that same image:</p>

<p> <a href="http://static.photo.net/attachments">http://static.photo.net/attachments</a>, etc.</p>

<p>and did not show any other images.</p>

<p>What am I doing wrong? Do you mean I should paste the html into the search box? Or do I first have to download the image onto my harddrive, and then upload it into google images? I guess this is an advanced function I just have not located on the google images page.</p>

[funcrunch]

<p>John - did you start the search from <a href="http://images.google.com/">http://images.google.com/ </a>? If so and it still didn't work, you might have Javascript disabled or an incompatible browser. It's even easier if you install Google's extension for Chrome or Firefox (<a href="http://www.google.com/insidesearch/searchbyimage.html">linked from here</a>); then you can just right-click and "Search Google with this image".</p>

[tom_mann1]

<p>BTW, a problem with Google's image search add-in is that it isn't yet compatible with Firefox 5.0.</p>

<p>Tom M</p>

[funcrunch]

<p>Tom - I'm using it with Firefox 5.0 on Ubuntu and it works fine for me...</p>

[tom_mann1]

<p>Hi Julie - I wish it did for me. Maybe the difference is because I'm running Firefox under Windoze (Vista). Here's what I get when I try ...</p>

<p>Tom M</p><div></div>

[dianedh]

<p>John,<br>

When I tried to drag my images to the search box, all that happened was my image opened. I ended up clicking on the camera image which opened a new box. Then it gave me the option of uploading an image, which I did and had no trouble then. I don't know why I couldn't drag either but I did get some interesting 'match' shots to the rose shot I actually posted!<br>

Diane</p>

[NetR]

<p>>If you want a really good laugh, just drag my image of two beetles mating (attached) into their search box<br>

What sort of images were you hoping to find, Tom?</p>

 

[tom_mann1]

<p>I hoped that their algorithm would be smart enough to classify the subject as "living critters", maybe even realize that there were two of them, maybe even be as specific as "two insects", and then at least return other images of insects, maybe even other types of similarly shaped and colored beetles.</p>

<p>I realize, of course, that performing such a classification is an extremely difficult task, given that the subject matter of images submitted to the search engine can be anything under the sun, at any magnification, any orientation, any number, on confusing backgrounds, with any lighting, etc. So, I was not terribly surprised that Google's Image Search took the vastly easier (and classic) approach of only attempting to match color, texture, perhaps rough placement in the frame, etc., and not even attempting to classify on the basis of content.</p>

<p>Based on a bunch of test images that I submitted to it, my guess is that they use an algorithm such as this: They first rescale all images to some common, quite small size (maybe even as small as 10x10), convert the resulting image to an HSL-like color space, quickly find the first few principle components (ie, http://en.wikipedia.org/wiki/Principal_component_analysis) of the L-channel, throw out the first component (which is average luminosity), and do similar analyses on the Hue and Saturation channels, again, discarding the 1st pple. component of each of these to get rid of differences in average hue (1st pple component of the hue channel), and average saturation (1st pple component of the saturation channel). It might also perform a conversion to polar coordinates to allow for tilted similar images to be found. Such an approach is well understood (ie, the rudiments are in every grad-level image processing textbook), mathematically simple, robust and requires minimal computing resources (fast - the length of the feature vector can be small), and is "safe" in that it is essentially guaranteed to always return "visually similar" images. Of course, the problem is that such an approach doesn't provide any insight whatsoever into attempting to classify the actual subject matter, so one gets matches which are pretty silly to humans.</p>

<p>For example, in the case of the test image I mentioned earlier, their algorithm felt that the 10 closest matches to a photo of two beetles mating on a white background were all photos of nicely prepared and presented orange colored food on a white background / tablecloth or plate (see attached).</p>

<p>;-)</p>

<p>Tom M</p>

<p> </p><div></div>

[MichaelChang]

<p>Tom, the new search function does give it additional refinement capability if the user provides an additional description.</p>

<p>If you type in "Beetles mating" in addition to your image search, you end up with this:<br /> [<a href="http://www.google.ca/search?hl=en&bih=1002&biw=1424&tbs=sbi%3AAMhZZivQp3w-vG5eU3zo8676_1kZ5lHlbiOW6G047BZC8SzecBUiXp9X1nWpSUBBtIH7w-UEDimY2I1jaVRV_162vlMrnXQ33t0yWExNPLtH7ufnRoTFwSYY5K8ufnIDUGDa0xlPcDkHnAyzcjFGXlnnrFp7iGgVu7P1yn2cMsDiDerqaTeVoKcdrfZRbEmZHI2TnayuWroftxZs92T1UwJNTKoH-A2ojA7DzfBnt55gD73uJZnEQ_1HaeX7J_1oqM-nn2P24_1pL0Nz8nP_1KIPNW5f2J7wezwOznF-2kX7BrkXmHR2ChjJPyoLLAw_1qwX7fhQjjqLuFpIJE2HSwTKKL4zRNxNoFpV0MpJR_11Nu66cacqwZ767ugyAk7MUx1moMRHgV1uPoMlBlIXAeLU7widwLPODgmT6wjlqlFoWspmM0GNCLZ2TTVtHL2X_1EHlS0CER6Bwwy055ddnWLrqs0FjK-K78Hfbb8tVALOAlX9tiWNEPNf5AZosHe273vgTrIfwvrSYJpzP7WDElpzjDxzh19wktZ1zH72BgLZDXA5it7H5dmyXfzMPC-5527AosWqaYzATJWuJp3vHVHdMZxmiL4W-IRk3xRPcBKL-LclOIejs5E2Q9kZDaiTz86pcxkjDL8o83VpFY-nPey8GWIi8cRWERQ5xOTRgqq5r-IYAYMc9nYxVoU_1Ww0ImR-uNcbFOjOPMF9hR1sGuEIyD8tzhD4JRQ5YpV9Sy2w-I-AARSJH71nVawCmMPrczr06p6-6Mw9wKM6P177RL9TSZvnywNJF7OPj1-N1MrzH8_1u2Pc5Nv5r-Tt4Q7CNtzYRAqWKP_1PR-AEgzXc9qOZx5yxY0z3x0aPgb9FwWwq60AKU-lUdqSW0Ekzi4c6QdFTlTnkar90aab_16iJ6pUwUqubAT1F3y6i-BA7AKMFgS5PPLqIelQKh5Xds7XITiQUmx_1001S9WvQzbUIKU3kFR6kmebqx1KAcqYoTuyG_1_1PVfR-nSLG6qzjiBtW94ImrPLLnK2fh6tvT_1FPCo1eSfv_1EjAH2G6rvvZKnOekBS97HoxNb1H68kwojXBLl4EJiVnglUwdfAqg2tu7t_1HQOg38YMJ4_1b1OZH_1SA551JfvtSvQJukz2l7pZCY9m6cvSfRfHdkUb_1LmpKw4Bl94kiC9sqRhb7QD6a9sfuJ598MXEnwdOUnetc40-2yKutKPMDaT1Zimn-CcF9xDTLB3ssNkmILJAJc57cy9vaTjlvv0ME2SP-0gkcmnxpVv8nJnxNJ9y4Gn3dDb63I8ffDjjW4bo_1t5-oXVP6zc80TdUS6Jw&q=beetles+mating&oq=beetles+mating&aq=f&aqi=g-s1g-v5g-m1g-bm1&aql=1&gs_sm=e&gs_upl=5164l8843l0l11006l16l15l1l0l0l0l217l2069l3.10.1l14">Link</a>]</p>

<p>I imagine using both will give you a substantially better chance of finding duplicates on the net (which was our original purpose).</p>

<p>PS: I changed the description to simply "Beetles" and the resulting matches were different as expected. </p>

[MichaelChang]

<p>I tried another search using the image below and used "Mona Lisa" as the description because the image sort of has that feel, and a good test since Mona Lisa seems to be the benchmark image used for Tineye.</p>

<p>The result was different from that without a description as expected, but using both seems to cover more ground than using either alone.</p>

<p><img src="http://d6d2h4gfvy8t8.cloudfront.net/3155222-lg.jpg" alt="" width="600" height="800" /></p>

[tom_mann1]

<p>Hi Michael - Thanks. I hadn't noticed that one could add text to an image similarity search. I thought it was an either/or thing - you either did an image similarity search or a text-based search for similar images. That's great! I'll have to play around with it.</p>

<p>Tom M</p>

[brandon_hebshih]

<p>i use the google search for images all the time</p>

<p> </p>

[jay_poel]

<p>Another good image seach engine is <a href="http://www.tineye.com">www.tineye.com</a> , it's been around for a while now.</p>