Human Visual System: Caucasians and Asians recognize faces and expressions differently

World Science (an online science news site) summarizes work by researchers at the University of Montreal (Canada) into how Asians and Caucasians differ in recognizing faces and facial expressions.  Eye tracking cameras were used to monitor where the subjects were gazing.  Doctoral candidate and researcher Caroline Blais said that

The study confirmed that Caucasians study the triangle of the eyes and mouth, while Asians focus on the nose..

 In a separate study, she also tested how recognition of facial expressions differed between the two ethnic groups. It turned out that Asians were not as accurate at recognizing emotions that required observing the mouth, such as fear, disgust, and anger.

The two studies were published in the journals Current Biology and PLoS One.

Image Recognition: Facial recognition feature in iPhoto

I installed Apple's iPhoto '09 last night to test how the face recognition feature (called "Faces", of all things) worked. There are nearly 4,000 photos in my personal library, so I left it running overnight while it went through the database and, I assume, ran its face detection algorithm. I then decided to run a test. I manually identified the same person in two different photographs and had iPhoto go out and locate matches, which is its next step in its learning process - the user next needs to confirm or not confirm the identity of each one of the identifications it produces.

So how did it do at the end of this first step? Surprisingly well is the answer, particularly for a consumer-grade product. The raw numbers were 72 correct and 16 incorrect at this first step, but that doesn't tell the whole story. In the instances where it was incorrect, all but three were confused with family members (which admittedly is the most likely confusion outcome since for all pictures with faces in the database, generally at least one is a family member). In addition, it also correctly chose pictures where only one eye was visible, where the age was significantly different, and where there were different emotions clearly visible on the faces. It also did not seem to have a problem with different compression levels, low resolution (including one that was very granulated), face paint, under-exposure, black-and-white, scans (of printed pictures), and color. Finally, it made only one error out of 86 pictures in correctly detecting and locating the face.

So, even though this was not a properly constructed test, the results were more than impressive enough to warrant looking at it further for professional use in some applications.

PS. This is not a product review, per se, but I should say that my experience with this just-released version was not all positive. Namely, it was horribly slow during the step where one has to select a picture, go into the Faces mode, and then type in the subject's name. Each entry took two to three minutes to accomplish, so I saw the "spinning beach ball" that Mac OS-X uses to let you know it is busy quite a lot. Hopefully this bug will get corrected and patched quickly.

Image Recognition: Facial and license plate recognition news items

There are two on-line news articles related to image recognition that I would like to bring to your attention.

The first news article is about law enforcement in Tacoma, Washington, USA using a facial recognition package to match 16 years worth of prisoner mug shots with pictures taken by ATM (Automated Teller Machines) in a forgery and theft investigation to generate the lead needed to solve the case.

The second news article is from New Orleans, Louisiana, USA where local law enforcement used a license-plate recognition system to make 20 arrests and recover 23 stolen vehicles and license plates in a 25 day period.

The common thread here is, of course, automatic image recognition. These software algorithms have come a long way in the last decade. However, one should understand that the conditions are partially or completely controlled in both of these applications - i.e. distance, lighting, exposure, aspect (turned toward the camera), and (possibly) compression all fall within acceptable boundaries. In addition, with the license plate recognition problem, the character set and fonts were known in advance. The controlled conditions and a priori knowledge significantly increase the accuracy of the results tremendously and the chances of a successful investigation and prosecution.

(Hat-tip to JUSTNETNews, USA - I highly recommend this free service of the National Law Enforcement and Corrections Technology Center, or NLECTC, which is part of the National Institute of Justice, USA)

Image Recognition: Military requirements push technology forward

The Strategy Page has an article about how the proliferation of video on the battlefield (e.g. from surveillance cameras and even night vision goggles) is driving technology development. Initially, it was the UK (United Kingdom) that drove image recognition technology with its massive deployment of CCTV in public spaces (e.g. train stations, airports, city centers, and shopping malls). Now, it is the military.

The benefit from computer-assistance in analyzing video comes from the following:

1. Volume (i.e. the sheer number of cameras and, hence, images to be monitored and/or analyzed)
2. Concentration (i.e. the human visual system loses the ability to concentrate effectively on images after about 20 minutes of continuous viewing)
3. Memory (i.e. computers can track and, possibly, predict more things simultaneously than a human can because, generally speaking, computers are not nearly as attention and working-memory limited in the short-term as humans and far outstrip us in their ability to recall video sequences over the long term - they can just play back the video recording off of their hard disks)
   

The article claims that the abilities of computer systems to recognize patterns is rapidly improving and approaching that of humans - that is no small feat as humans are simply amazing in their ability to perform real-time pattern analysis.

One intriguing point made in the article is that the current conflicts are generating a lot of real-world recordings of "bad behavior" that can be used to train and test new pattern analysis and prediction algorithms - training data like this is like gold to us signal processing types!

Enjoy!

Image Recognition: Military uses pattern recognition to detect bomb planting activity

I wanted to call your attention to an article by the Strategy Page on how the US military in Iraq is using pattern recognition on imagery to detect IED (Improvised Explosive Device) emplacement activities.

A very simple implementation of this technique is to take a picture on Day 1, then another picture from the same location and in the same direction on Day 2. Next subtract the two images from each other and see what is left over. All the things that did not change between the two images will disappear and only those things that changed will be left! In this case, what will be see on the "difference image" might be some disturbed gravel, tire tracks off the side of the road (assuming it is a road-side bomb emplacement) or, if the military guys are really lucky, what will be left is a bunch of guys standing around with shovels and a bomb as they work to put it in place. Neat (and helpful)!

Image Recognition: "Computer, where did I lay the keys?"

The Daily Mail (UK tabloid newspaper) has a gadget article about Smart Goggles, a human wearable image recognition and recall system built into a set of glasses. The user trains the system by focusing its built-in camera on a series of objects, such as car keys, CDs, etc. while speaking the appropriate name of the object. Once trained sufficiently, the system automatically recognizes images it sees and stores the information for later retrieval, again by spoken word. The processor (computer) is worn on the user's back. The system was developed by Professor Kuniyoshi and colleagues at the University of Tokyo.

It isn't much of a stretch to see how these same concepts could be built into other video applications, such as CCTV systems, for instance. Don't be surprised if they don't work this into a future episode of CSI or similar television show (or the next James Bond movie, for that matter).

Image Recognition: License plate scanners

The technology manufacturing cycle is at work with license plate (tag) scanners - decreasing prices lead to more sales and deployments, which lead to further cost decreases and so on.  We've all seen it with personal computers, cell/mobile phones, DVD players, flat screen displays, and a multitude of other devices.  Now this market force is at work with license plate scanners , which are now being deployed on police cars.  The scanners are automating what was before a completely manual process - namely running stolen tag numbers and such - and doing it much faster.