Which technique is the best for facial recognition?

I have to apply a facial recognition technique on my project. I have read many research papers but i couldn't finalize the best technique including Hidden Markov Model, Support Vector Machine and Neural Network for my scenario.

Can any body tell me which will be the best for my case?

and on what basis do we decide the suitable technique?

[Mehrukh Kamal]

I don't have much time to check all of them. My case is Facial recognition in TV broadcast.

Voila Jones is for detection not recognition and i m already using MATLAB's voila jones for detection of faces.

As far as acurracy is concerned there is not any level of requirement. but yes obviously it should be high enough to prove my chosen technique.

[Walter Roberson]

We don't have time to read all the research papers and do a detailed theoretical analysis for you and report back to you so you can then implement the "best" technique.

Facial recognition in TV broadcast: so your program needs to be able to track individual soccer players as they run, by recognizing the faces? Or it needs to be able to tell which national politician is giving a particular press conference ?

[Mehrukh Kamal]

national politician is giving a particular press conference

The research papers that i have read have just written that they have used this technique they havnt mentioned that why did they used this technique.

Answer 1

[Walter Roberson]

No, we cannot tell you which will be the best for your case, as you have not indicate what your case is. Are you trying to recognize individual shoppers as they go through a store? Are you trying to recognize known casino cheats? Are you trying to take automated class attendance? Are you trying to verify that an astronaut has not been infected with an alien parasite before allowing the astronaut access to the control room with the Illudium Q-36 Explosive Space Modulator?

Answer 2

[Image Analyst]

I bet each one of those papers says that their method is the best algorithm for their facial images, right? Those methods you mentioned don't sound like face detectors I have heard of but maybe they are used in some steps of the algorithms in the papers you read. The algorithm I hear of most often , and is used by MATLAB's Computer Vision System Toolbox, is the Viola-Jones method. It uses Haar wavelets. I'm sure there have been improvements upon it since it was introduced.

The basis upon which you decide if it is suitable is if it detects faces to the accuracy you require. If you have several algorithms to choose from, you should construct an ROC curve for each one to help you pick the best one. https://en.wikipedia.org/wiki/Roc_curve

[Walter Roberson]

Not just "detects faces to the accuracy you require" but also "detects faces under the circumstances you require". Training on full-on passport pictures is not likely to have good results on pictures of people in parkas on perambulating parking patrols.

[Greg Heath]

Confucious:

Try all; Choose best.

[Mehrukh Kamal]

isnt Voila Jones for face detection? (not recognition)

[Image Analyst]

True. It will detect a face. To detect exactly what individual's face it is, you're going to have to measure a lot more features.

[Mehrukh Kamal]

The Voila Jones code present in the matlab gives me the face eyes nose and mouth. arnt they enough if i do pattern recognition from them?

[Image Analyst]

No. Everyone has eyes, nose, and a mouth. Like I said, to determine if it's my mouth or your mouth, you're going to need higher resolution than needed for detection and better descriptors regarding shapes of the features and relative spacings. Check out this page for algorithms: http://iris.usc.edu/Vision-Notes/bibliography/contentspeople.html#Face%20Recognition,%20Detection,%20Tracking,%20Gesture%20Recognition,%20Fingerprints,%20Biometrics

[Mehrukh Kamal]

okay.. can you tell me the criteria for choosing a particular machine language?

[Image Analyst]

You can construct an ROC curve to help with that: https://en.wikipedia.org/wiki/Roc_curve

Answer 3

[Mehrukh Kamal]

I don't have much time to check all of them. My case is Facial recognition in TV broadcast.

Voila Jones is for detection not recognition and i m already using MATLAB's voila jones for detection of faces.

As far as acurracy is concerned there is not any level of requirement. but yes obviously it should be high enough to prove my chosen technique.

[Jeremy Wurbs]

Viola-Jones is a method for training a classifier to detect a certain object. The original paper trained a classifier to detect faces, and is one of the most common methods for training such a classifier today. That said, you could also train a classifier for particular user's faces and get a classifier for each of them.

The techniques you mentioned in your OP (SVM, NN, etc.) are general methods for taking labeled data, X, with labels, Y, and training a classifier, f(x), to predict the output class on new, unlabeled data. Such techniques fall into the realm of machine learning. In theory they will all work, but for vision applications the difficulty is almost always in finding meaningful features (X) and handling image pre-processing (centering the face, removing unwanted edges, etc.). These are, in general, very non-trivial tasks.

[Mehrukh Kamal]

So whats the criteria for choosing a perticular machine language? and where can i find it?

[Walter Roberson]

The criteria for choosing a particular algorithm is that it works "better" than the others under conditions that are realistic for what you need. "better" might mean less false positives, fewer false negatives, time required, expense required, and so on.

There was a presentation on tracking names and voices in TV broadcasts in ISDL 1997 http://www.dl.slis.tsukuba.ac.jp/ISDL97/exhibition-en.html