BACKGROUND OF FACE RECOGNITION

Face recognition under uncontrolled image acquisition conditions is a challenging goal, not only because of the gross similarity of all faces, but also because of the vast differences between face images of the same person due to variations in lighting conditions, facial expression, and pose.  shows example images of the same person with the lighting source positioned in front, above, bottom, right, and left of the person, respectively. shows example images of a person with neutral, smiling, surprising, and sad expressions, respectively. Finally shows example images of the same person with different presentation angles.

Training of Frontal Face Models

To train the frontal face model, we collected face image samples from 40 individuals for the training set. For each person, we had three I’m ages in three pose views (left 15°, frontal, right 15°); all these images are extracted from the Ferret database. We labeled each of these 120 face images with 58 points around the main features, including eyes, mouth, nose, eyebrows, and chin. Some labeled training face images can be seen in.

REAL-TIME AUTOMATED FACE RECOGNITION SYSTEMS

To date, the majority of the research work on automated face recognition (AFR) has focused primarily on developing novel algorithms and/or improving the efficiency and accuracy of existing algorithms. As a result, most solutions developed (similar to the examples given in previous sections) are typically high-level software programs targeted for general-purpose processors that are expensive and usually non real-time solutions. Since face recognition is typically the first step and frequently a bottleneck in most solutions due to the large search space and computationally intensive operations, it is reasonable to suggest an embedded implementation specifically optimized to detect faces and recognize them. An embedded solution would entail many advantages such as cost and miniaturization, as only a subset of the hardware components are required compared to the general computer-based solutions. The resulting solution can then be integrated with other technologies such as security cameras to create smart devices.

Now that reliable, accurate, and efficient face recognition algorithms are available, coupled with advances in embedded technologies, low-cost implementations of robust real-time face detectors can be explored. The following subsection discusses the common embedded technologies and known embedded implementations of automatic face recognition (AFR) systems.