Sometime over the summer, and apparently without enough fanfare for it to ping my sonar, Google announced a new “search by image” service. The idea is that you can either upload an image file, or post a link to an online image, an Google will go out and search either for matches to that image, images that are “visually similar”, or if the image can be identified, links to information about that image. I finally got a few spare minutes to take a look, so I decided to try a few experiments.
Only recently did this new service come to my attention. My fellow Flickr photographer Eric Morris, aka The Rested Traveler, had posted a link on his Facebook page to a blog post where the writer used the image search to see if his images had been used without his permission. Having been the victim of photo theft before, I thought this was a good starting point.
I first started with an image I knew had been reposted. This photo gets tons of hits, and I already knew it had been posted elsewhere:
If you’re doing this with a Flickr photo, you first have to go to the page where you can view/download all sizes.
It doesn’t matter which size you pick. Just right-click on the Download link and select Copy Image Location (Firefox). Go to Google and paste the URL for the image into the search box. In this case I get the following message:
In this case, select the “search by image” link. The search results first showed exact copies of my image. I could tell that they hadn’t hot-linked to my Flickr photo, but had reposted the image. The addresses were the give-away.
All of these were hosted elsewhere. Just goes to show that Flickr’s security protocols aren’t as tight as they would like.
I was curious as to whether or not the search would do facial recognition. I searched based on a photo of myself from an ID card. The results were interesting…
These were all photos using a similar background. However, if I used a photo that has a more intricate background, such as a photo of me in Washington State, the results weren’t as predictable. It did, however, find that I had used this as a profile picture on Facebook.
I was curious about the image recognition software, and whether or not it would recognize common icons or landmarks. It does, but with some limitations. I searched on a photo I had of the McDonald’s arches, and it didn’t recognize it at all. I also searched using an evening photo I had taken of the Lincoln Memorial. Here were the results…
All it returned were dark-ish images in a similar color palette. I found that I could, however, give it a hint and get better results. Here I’ve entered the word “Lincoln” and gotten better results…
I knew that some smartphone apps let you snap a picture of a DVD or CD cover, or the cover of a book, and will return results from Amazon.com and other sources for that publication. I wondered if the image search would do the same thing. I’ve been reading Walter Edgar’s “History of South Carolina”, so that was my starting point. I took a fairly close-up shot so that only the book cover is seen, and it recognized the cover.
It apparently did some OCR on the image, because it read the text on the cover. Further down in the search results it also returned links to a Google Books preview of that book. I’m just curious as to what would be returned if I snapped a photo of the book cover on a cluttered desk. I may have to give that a try next. I also wonder how complex the OCR system really is, and how much text can be read for a search.
So, it appears that the new search by image has some sophisticated algorithms in place, but results aren’t perfect. You can help the search along with keywords. It’s great for finding stolen images, and it can be used to find images in the same color scheme. I don’t know how much I will use this, but so far it’s been very cool.