PhD project: Automatic Face Recognition by using Machine learning Algorithms

Name: M.F. (Faik) Karaaba

Supervisors:
prof. dr. L.R.B. (Lambert) Schomaker
dr. M.A. (Marco) Wiering

Summary of PhD project:

The main objective of this PhD project is to automatically detect faces given a photograph and then process and finally recognize. Faces can be considered non-rigid and non-fixed objects so that a face can appear in a photograph in a very different angles and statues. Now the steps for a face recognition is explained below:

1)     Face Detection

The face, which is to be recognized, should be detected first. Face detection is basically a binary classification that is carried out in several image patches which can contain a face. These image patches are usually obtained in a sliding window method. While there are some commercial and/or well-known face detection packages (the most famous is Viola-Jones face detector) , there are also several machine learning algorithms for those who want to build their own. We used in this project Viola-Jones face detector.

2)     Face Alignment

After the face is automatically cropped, it should be normalized in order to avoid rotational noises. To do this, in our project, we use eye coordinates to normalize a face rotationally. Then, a good alignment (which also affects the recognition performance) depends on how accurate centers of the eyes are located. Here, similar to face detection, we use a sliding window to detect eyes. However, differently from the other eye detectors which search through the face, we first locate the eye-pair region and the locate the eyes. In the following diagram this process is depicted.

3)     Face Recognition

Face recognition has two meanings: The first is deciding of whether given two faces belong to the same person or not. It is specifically called face verification. Second is that, given the face of a person, the task is finding the identity of this person, by searching in a face database. It is called face identification.

We focus on face identification in this PhD research.

4)     Data Preparation - Feature Extraction

Although image files can be used directly for training a classifier, this is generally a not a good approach, because there can be variety of noises which possibly challenge the classifier's job depending on the resolution of the image of an object to be detected/recognized. Therefore, input image data is often first preprocessed by an image filter (e.g. an edge detector). This process is called feature extraction. Histogram of Oriented Gradients (HOG), Scale Invariant Feature Transform (SIFT), Restricted Boltzmann Machines(RBM) are some of the well known feature extraction methods/algorithms.