Face recognition
Download as PPT, PDF•1 like•380 views
This document summarizes face recognition techniques. It discusses three levels of facial details, from gross to micro features, and how they require different image resolutions. It also outlines the major components of a face recognition system: image acquisition, face detection, and face matching. Finally, it describes common image formats like 2D photos and 3D scans, and detection methods like Viola-Jones that use Haar-like features and AdaBoost training.
Face recognition
- 2. Face recognition can be defined as the process of establishing a person’s identity based on their facial characteristics. face can be captured at a longer standoff distance using non-contact sensors. Training a machine to recognize face images exhibiting such unconstrained intrauser variations is a difficult task
- 3. The underlying mechanism of face perception in humans has been studied for two purposes: (a) to design machine recognition systems that can mimic the human ability to recognize faces (b) to understand the neurological or psychological mechanism of brain functions for medical treatment. Studies using advanced brain imaging techniques such as functional magnetic resonance imaging (fMRI) are expected to reveal the precise face processing mechanism in the human brain.
- 4. the facial characteristics can be organized into the following three levels: Level 1 details consist of gross facial characteristics that are easily observable. Examples include the general geometry of the face and global skin color. These features can be extracted even from low resolution face images < 30 interpupilary distance (IPD)
- 5. Level 2 details consist of localized face information such as the structure of the face components (e.g., eyes), the relationship between facial components and the precise shape of the face. they require a higher resolution face image (30 to 75 IPD)
- 6. Level 3 details consist of unstructured, micro level features on the face, which includes scars, freckles, skin discoloration, and moles. Problem : Level 3 details may be critical is the discrimination of identical twins
- 7. A typical face recognition system is composed of three modules: (a) image acquisition, (b) face detection (c) face matching
- 9. Automated face recognition requires the face data to be in a machine readable format. Conventional 2D photographs, 3D range or depth images, and videos are the three major types of image formats used in face recognition systems.
- 10. In general, the frontal view of a face contains more details than a profile view Recognition systems based on 2D face images are also greatly affected by variations in illumination and spatial resolution. To overcome these challenges, new sensors such as high resolution cameras, active pan- tilt-zoom (PTZ) cameras, and infrared cameras are being used.
- 12. A NIR camera can operate even under low illumination conditions because it uses a separate NIR illuminator. A PTZ camera can dynamically zoom in or zoom out to obtain close-up images of objects of interest. Multi-camera configurations that capture face images at multiple pose angles have been used to address the pose variation problem.
- 13. The inherent pose, expression, and lighting variation problems in 2D face images stem from the 2D rendering of a 3D face object. There are two types of 3D face capture systems: laser scanning stereographic reconstruction.
- 14. laser scanners provide more accurate 3D face models. stereographic cameras provide near real-time capture capability with slight loss in accuracy. The image captured by a 3D sensor typically covers about 120◦ of the human head and this image is referred to as a 2.5D scan.
- 15. 3D face models are usually represented as a polygonal mesh structure (e.g., triangular or rectangular) for computational efficiency. Thedrawbacks of 3D face acquisition include longer image acquisition time, the large data size of the 3D model (which requires higher computational resources during matching), and the relatively high price of 3D imaging sensors.
- 16. A video camera can continuously capture face images, thereby enabling the selection of a good quality face image (e.g., one with frontal pose and neutral expression) for recognition. Face recognition in video has attracted interest due to the widespread deployment of surveillance cameras.
- 17. Two distinctive pieces of information are provided by a video stream: (a)multiple frames of the same subject (b) temporal information (Dynamic Facial Motion) pertaining to an individual’s face.
- 18. Face detection methods are typically based on extracting local texture features from the given image and applying a binary (two-class) classifier to distinguish between a face and non-face.
- 19. The Viola-Jones face detector scans through the input image with detection windows of different sizes and decides whether each window contains a face or not. The feature values are obtained by computing the difference between the sum of the pixel intensities in the light and dark rectangular regions.
- 20. 2D rectangular filters are similar to the one- dimensional Haar wavelets used in the signal processing domain, they are also known as Haar like filters. No single Haar-like filter can perform the face detection task with high accuracy. For instance, more than 180,000 different Haar-like featurescan be derived from a detection window of size 24× 24 pixels.
- 21. The computational burden involved in the computation of these features can be significantly reduced by pre-computing the integral image from the original image. An integral image S associated with the given image I can be computed as follows:
- 22. where I(x,y) is the intensity value of pixel (x,y) in the original image and S(x,y) is the value of the corresponding pixel in the integral image. Though the integral image allows efficient computation of the feature values, it does not reduce the computational complexity to a level that is sufficient for realtime face detection.
- 23. A variant of the Adaboost algorithm proposed in the pattern recognition literature can be used to select the discriminative features as well as to train the classifier function. The feature selection and classifier training algorithm works as follows. Let f j(w) be the feature value obtained by applying a filter f j to the detection window w, j = 1,2,··· ,L, where L is the total number of filters. The detection window w isparameterized by its location and size. Each feature value can be used to constructa weak classifier that predicts whether the window w is a face image.
- 24. Weak classifier can be defined as where p j is the sign (+ or -) determining the direction of the inequality, θ j is the threshold, and j = 1,2,··· ,L.
- 25. For example, an eye detector can be constructed by simply replacing the training set with examples of eye images and non-eye images. the Viola-Jones face detector is not perfect and can produce both false positive and false negative errors. A false positive error refers to the detection of a face where none exists, while a false negative error indicates that a face present in the image was not detected.