Description:Field of Invention
Face recognition is a task that people perform naturally and easily in their daily lives. The high availability of powerful and as well as embedded computing systems have attracted a great deal of interest in the automatic processing of digital images and videos within many applications. Automatic face recognition has turned out to be one of the most interesting and essential tasks in the field of computer vision and biometrics, concerning theoretical techniques and software to recognize individuals in light of their facial pictures. The importance of face recognition is due to the many applications that can be encountered in different areas such as: security, monitoring devices in public spaces, forensic applications, identity of people in image or video databases, human-machine interface, smart cards and the biometric passport also known as E-Passport. The field of facial recognition exploits the knowledge of many disciplines such as: image processing, applied mathematics, pattern recognition, machine learning, visual perception, psychophysics and neuroscience.
Background of the Invention
Face recognition is the process of identification of humans by faces. The main aim of recognizing the face is identification of individual from similar class of species. The individuals are identified by using unique features of the face. The automatic facial recognition is demanded not only to be used by the security forces it is important to control fraud, verifying the identity of clients. For example in accesses to bank ATMs, to areas of restricted access, to multimedia telecommunications, in video surveillance systems, etc.
The majority of face recognition processing algorithms are based on vector or matrix data models. These operations destroy the data structure and may result in decreased performance and/or robustness of processing in various applications. It is therefore necessary to adapt the conventional treatments for these new configurations (multidimensional). Multi linear algebra makes it possible to exploit these data while preserving their structure. The data is then represented as multidimensional tables called tensors. More recently, these techniques have also been extended to facial biometric data. However, the generalization of facial data processing in 2 conventional vector or matrix form to the tensor case is not yet obvious (US9639740B2).
During the last twenty years, the automatic recognition of faces has a key issue, particularly in the areas of indexing of multimedia especially in security. The main problem in the real world is detection of faces with high performance. The conditions like occultations, illumination variations, pose variations and facial expressions plays a crucial role for achieving the performance of facial recognition. And it mainly depends on lighting condition the color images does not predict the face as an object. Different variations of the poses also plays a key role for face recognition. To overcome this type of problems pose correction and rotation angles of the face are necessary. To develop this faces recognition using pose variations leads to the problem of increase the cost in terms of time, memory. Another issue for recognizing the faces automatically is facial images representations and extracting the most important features form the face. For extracting local and global features many number of methods are developed. For representation of images local descriptors are mostly used (US9721148B2). To improve the performance for facial biometric the research is mainly focused on use of 3D surfaces and shape of the face for describing the most important features by increasing the number of dimensions.
Summary of the Invention
The main goal of the first step of system is the detection and localization of the facial area in a given image. After the detection step, the biometric signatures of human face are extracted as a characteristic vector within the second step. After formulating the representation of face image, the final step is reorganization of the individuals. For every individual, several pictures are taken and then extract the characteristics and finally stored in database (off-line). Then, when an input face picture arrives, perform face detection, feature extraction and finally comparison of these features to each face class stored in the database (on-line phase). The main objective of this invention is to develop, implement and evaluate a new automatic face recognition system based on images, in which the variations of illuminations, expressions and poses are very different between the learning set and the test set. In particular, a new hybrid method of an effective strategy for local and global features representations and reduction of dimensionality for automatic face recognition.

Brief Description of Drawings
Figure 1: The process of recognizing a face.
Figure 2: Block diagram of Holistic hybrid face recognition system
Figure 3: Block diagram of face recognition system for Fusion.
Figure 4: Block diagram of Hybrid face recognition system.

Detailed Description of the Invention
Facial recognition is identification of faces in an image or video automatically. The system can operate in the following two modes: authentication or identification. It can also note that there is another type of facial recognition scenario putting at stake a check on a watch list, where an individual is compared to a list restricted number of suspects. The basic operating principle of a recognition system can be summarized in three steps when standard data set is acquired for recognition model and it is shown in Figure 1. First step in the face recognition detects (or tracking) a person's face in an image or video, aligns the face and allows normalization of the detected face image, so that it can be compared with other face images on the dataset that do not necessarily have the same size or the same lighting. Therefore, this requires the detection of certain features of the face (usually the eyes, nose and mouth) and the normalization of the existing distances between these features. It also involves knowing the pose of the face. The illumination normalization step also requires a particular treatment. Second step is Extraction of facial features. Once the alignment is done, it is necessary to extract the most relevant characteristics of a face to allow its identification and not to confuse it with that of another individual. The feature vector thus associated with a face must be robust to all possible variations, such as expression, pose and illumination. Third step is the classification step: This step is a decision step that determines the identity of a person by comparing feature vectors.
Face detection can be done by detecting the color of the skin, the shape of the head or by methods detecting the different characteristics of the face. This step is more delicate, when the acquired image contains several objects of face or a non uniform background that creates a texture disrupting the proper segmentation of the face. This step is dependent on the quality of the images acquired. In the literature scientifically, the problem of localization of faces is also referred to by the terminology "face detection". The overall performance of any automatic system recognition largely depends on the performance of face detection. In the detection step, identification and location of the face in the image is acquired at the beginning, regardless of position, scale, orientation and lighting. Detection approaches can be divided into four categories: methods based on knowledge, where is codes the human knowledge of the face, methods of correspondence masks, methods with invariable characteristics where color, textures and the contours and finally the most prevalent methods and which are those based on learning or statistics like PCA, SVM, Graph matching.
In the physical world, there are three parameters to consider: lighting, posture variation and scale. The variety of one of these three parameters may lead to a distance between two pictures of similar person, greater than that separating two images of two different individuals. The role of this step is to eliminate the pests caused the quality of the optical or electronic devices when acquiring the image in entry, in order to keep only the essential information and thus prepare the image for the next step. It is essential because you can never have a picture without noise because of the background and light that is generally unknown. There are many types of image quality processing and enhancement, such as: standardization, histogram equalization, filtering and gamma correction
The proposed method for automatic face recognition is shown Figure 2, includes two approaches for recognition of facial images. First the image is filtered using DWT, 2D-PCA and DWT, ICA then features are classified using PCA-ICA-NN, DWT-PCA-NN, and DWT-PCA-ICA-NN algorithms.
The wavelet transform is used to decompose low frequency images so as to differentiate high frequency components, in view of its capacity to catch particular transformed information of extracted image. The arrangement of the data into multi resolution frequency permits to confine the frequency segments acquainted by intrinsic values with expression or extraneous components (i.e. light) into several sub bands. These techniques cut away these different sub bands and spotlight on the sub bands which contain the most applicable data. he filtration and recognition using ICA and PCA are required but due to performance drop in case of occlusion, illumination, pose and expression etc. the discreet wavelet transform is used.
The face Face Recognition Using DWT-PCA-ICA-NN Algorithm involves the following steps. For Pre-processing the steps are : 1. Take image I1 and I2 as an input. 2. Resize both images for 160×160 matrix. 3. Apply RGB to Gray value conversion. 4. Histogram Equalization is performed on the gray image for Image enhancement. After completion of preprocessing next step is Feature Extraction. The steps involved are 1. After pre-processing, all the images are saved into database. 2. Convolve the images with the LPF and HPF filters 3. The images are decomposed into low frequency and high frequency components. 4. The low frequency components are down sampled by the factor of 2 in order to keep the even indexed columns. We get the approximation coefficients as its output 5. The high frequency components are down sampled by the factor of 2 in order to keep the even indexed rows. We get the detailed coefficients as its output 6. To obtain Level-2 decomposed coefficients, the rows and the columns of entry are further convolved with the filter. And the horizontal, vertical, diagonal features are extracted. 7. Compute the mean value of four coefficients (LL, LH, HL, HH) for the selected window 8. LL sub-band co-ordinates is again considered for reference by PCA measurement for each sub-window 9. Each sub-window is changed into a row vector Vi (i =1, 2 …... n) 10. The standard deviation σ and mean μ of vector components (W) is computed PCA matrix coefficient is computed by PCA measurement on A 13. The decomposed coefficients of image I1 and I2 are fused on the basis of PCA algorithm by extracting Eigen value and Eigen vectors. 14. From Eigen values and Eigen Vectors, independent component analysis (ICA) is calculated with the help of whitening and learning. 15. Calculate centre for test images with database. 16. Multiply centred information of test image to Eigen vector of database, 17. Perform whitening on output. 18. Calculate cosine distance between the feature of reference and features of test image 19. Threshold is given by cosine distance and finally result is taken out. 20. I-DWT is applied in order to achieve final fused image ‘F’ 21. The performance matrix is evaluated using reference image ‘I1’ and fused image ‘F’.
Principal Component Analysis (PCA) is a very well-known feature extraction technique used by face/pattern recognition systems. A new feature extraction technique known as two-dimensional principal component analysis (2D PCA) has been developed for image projection. The 2D-PCA is based on two dimensional matrix i.e. it need not to be changed into vectors. On the other hand, the covariance matrix can be simply generated by using the original image matrix parameters. If compared, the dimension of covariance matrix generated by 2D-PCA is much smaller than the covariance matrix generated by 1D-PCA. 2D-PCA provides proper feature extraction with lesser computational difficulty.
ICA is a statistical feature projection technique. This technique is the further extension of PCA. ICA projects the image data from high-dimension to low dimension. Fast-ICA method finds maximum of Non-Gaussianity of whitened data in order to compute independent components. The independent components (ICs) are actually the basic building blocks for representing the two-dimensional images. This fundamental idea behind ICA makes it useful in face recognition. Till now various ICA algorithms has been introduced. Some of them use adaptive scheme based on stochastic gradient methods. And the others find the ICs through minimization or maximization of high-order cumulates.
There are few limitations of global features to overcome those issues local features are used to claim higher accuracy with lesser training overhead in automatic face recognition system. In this the feature extraction methods DWT, PCA and point based features (Harris corner and SURF) are considered to develop a fusion based system for automatic face recognition. According to Figure 3 after face image database acquisition pre-processing performs and face detection is achieved with Viola–Jones method. Further detected face region is sampled for feature extraction, once feature is extracted with DWT+PCA, Harris corner and SURF, executed separately, as a outcome there will be 3 ID’s achieved respectively. All 3 ID’s are further fused to get the final ID. To overcome the limitation of trail and hit concept of fusion based system an efficient hybrid method which can elaborates the local features (LBP, HOG, Gabor wavelet) and global features (DWT, PCA) of face are implemented.
Hybrid methods are approaches that use both global and local characteristics of a face image. The performance of these methods includes the choice of the combination and how to combine them in such a way that their advantages are preserved and their disadvantages are avoided. These problems are similar to those multiple classifier systems or learning packages in the field of automatic learning. Local characteristics and overall characteristics will provide additional information relevant to the classification. The Hybrid approach for face recognition is shown in Figure 4. The features of face are extracted through conventional DWT, LBP+HOG and Gabor Wavelet and GLCM approaches and further these feature vectors are sampled through PCA with reduced dimension to Neural Network for classification.

4 Claims & 4 Figures , Claims:The scope of the invention is defined by the following claims:

Claim:
1. Automatic Face Recognition Based On Local And Global Features using hybrid Machine learning algorithms comprising the steps of
a) Examines and surveys the present use of Machine learning approaches for automatic face recognition.
b) Adapted different feature extraction techniques and propose a hybrid global feature based face recognition system.
c) Presented different techniques for Enhancing feature extraction and automatic face recognition system using Fusion based decision.
2. Automatic Face Recognition Based On Local And Global Features using hybrid Machine learning algorithms as claimed in claim 1, hybrid global feature based face recognition system, where feature extraction is done with Discrete Wavelet Transform (DWT), Independent Component Analysis (ICA) and Principal Component Analysis (PCA) methods. Further this method is classified with Neural Network (NN) and Support Vector Machine (SVM).
3. Automatic Face Recognition Based On Local And Global Features using hybrid Machine learning algorithms as claimed in claim 1, a novel face recognition system using Fusion based decision according to different ID generated from Seeded Up Robust Features (SURF), HARRIS corner and DWT+PCA feature extraction techniques for higher accuracy.
4. Automatic Face Recognition Based On Local and Global Features using hybrid Machine learning algorithms as claimed in claim 1, a novel hybrid face recognition system using DWT, Histogram Oriented Gradients (HOG), Linear Binary Pattern (LBP) and Gray Level Co Occurrence Matrix (GLCM) features with dimension reduction using PCA for lesser training overhead in classifier and then implement a real time face recognition system using Raspberry Pi Processor.