Registers are the fastest memory on the GPU, therefore it is important to limit the number of local variables for high performance applications. Shared memory is a read/write memory that allows memory sharing across threads in the same block. This is achieved by local DRAM blocks within each streaming multiprocessor. Overall this allows for much faster data storage without contributing to the global memory bandwidth. Global memory is a read/write memory that allows memory access by any thread. All incoming data and outgoing results must pass through global memory and as a result it is often the main bottleneck of the GPU. Another form of memory is constant memory which only allows read operations during runtime. Like global memory constant memory can be accessed by any thread, however it is highly cached making it much quicker than global memory. The final memory type is texture memory which is a special read only memory that has been optimized for texture based operations. Using special hardware built into the pipeline several common texture functions can be performed automatically including pixel interpolation and border wrapping. Unlike other forms of memory on the GPU, texture memory has constant access times for both cache hits and misses which allows for better scheduling and a 2D cache which gives it greater 2D spatial access.
Feature Extraction is achieved with the help morphological hit miss transform in order to detect face in the image. Morphological operations are useful for extracting image components that are useful in the representation and description of region shape, such as boundaries, skeletons, and the convex hull, etc. The hit-and- miss operation is a morphological shape detector that can be used to look for particular patterns of foreground and background pixels on an image. The hit miss operation is another form of dilation-erosion-based convolution. It is nothing but difference between erosion and dilation. Grayscale image is applied as input to hit miss transform. Firstly, morphological Erosion is done on this grayscale image. The basic effect of the Erosion is to erode away the boundaries of regions of foreground pixels. Fig. 3 (a) shows Eroded image. Again from the same grayscale image morphological dilation is done. Dilation is nothing but to gradually enlarge the boundaries of regions of foreground pixels. Fig. 3 (b) shows dilated image. Morphological techniques typically probe an image with a small template known as a structuring element. In this work, we are using one 5×5 disk shape structuring element. The care must taken, when we take a difference between Erosion and Dilation. It must be absolute difference instead of simple difference. Fig. 3 (c) shows hit-miss operated image.
Next, we obtain a surveillance video of the entrance lobby at INRIA labs in France from the same CAVIAR database . We keep only the first N = 295 frames, vectorize them, and arrange them as columns of data matrix X ∈ R 54338×295 . We set n = 14 and τ = 0.9 and repeat the experiment to obtain the background and foreground of the 80-th frame of the video using K = 3 PCs. In Figure 4.13, we plot the performance of ISVD , GRASTA , PCP , OR-PCA , RPCA  and ReProCS  (background and extracted foreground). We observe that ISVD , GRASTA , PCP , and RPCA  demonstrate again similar performance as before –i.e., ghostly appearance of the man in the extracted background and his blurred figure in the foreground. The method of  extracts a clean background but "ghostly" appearances of the foreground movement in the last frame is seen on the extracted foreground. On the other hand, OR-PCA , ReProCS , and L1-IPCA obtain similarly clean background and well-defined foreground.
ABSTRACT: This paper presents the two novel face detection techniques which are based on the singular vector decomposition (SVD) and Eigen value decomposition (EVD). Here, PrincipalComponentAnalysis (PCA) and Linear Discriminant Analysis (LDA) methods are applied to detect the features of faces which act as the principle component for the face recognition problem. The human face is full of information but working with all the information is time consuming and less efficient. It is better get unique and important information and discards other useless information in order to make system efficient. Principalcomponentanalysis is applied to find the aspects of face which are important for identification. Eigenvectors and eigen faces are calculated from the initial face image set. New faces are projected onto the space expanded by eigen faces and represented by weighted sum of the eigen faces. These weights are used to identify the faces. To reduce the time complexity and Euclidean distance in face space, here two techniques Singular Value Decomposition and Eigen-value decomposition are utilized. Simulation results have been presented to illustrates the effectiveness of the proposed face detection techniques. Here, face detection is performed usingPrincipalComponentAnalysis and Linear Discriminant Analysis methods with Singular Value and Eigen-value decomposition. Experiments on face database shows the effectiveness of our proposed algorithm and results compared to PCA using Eigen Value Decomposition(EVD) shows that the proposed scheme gives comparatively better results than previous methods in terms of reduced time complexity without effecting its accuracy. Based on the results presented it is concluded that the PrincipalComponentAnalysisusing EVD leads to superior results compared to that of other reported methods. From the result it is also noticed that the time complexity of the proposed method reduces considerable which leads to real time applicability of the proposed method.
Evaluating the performance of biometric algorithms is a difficult issue. Currently, there is also no detailed comparison among the iris recognition methods . For the purpose of comparison; we implement these methods according to the published papers. To compare their performance, we construct an iris image database named CASIA Iris Database. We use images of eyes from 30 persons, and every person has 10 images of eyes. MATLAB image processing tools were used to implement system. We use the usual methods to locate and normalize iris regions, and use the three methods mentioned above to extract the feature. Therefore, we only analyze and compare the accuracy and computational complexity of feature representation and matching of these methods. For each iris pattern, we randomly choose several samples for training and the rest for testing. After feature extraction, an iris image is represented as a feature vector. We used Euclidean Distance and Hamming Distance similarity measures to measure the similarity of iris features. Two distance measure lead to similar results and recognition result does not vary drastically. Recognition rate is shown in Table 1
Lastly, the CK+ dataset was incorporated to assess the joint face and expression recognition capabilities of various algorithms. Each test image is sparsely coded via a dic- tionary of both identities and universal expressions (Anger, Disgust, Fear, Happiness, Sadness and Surprise). The least resulting reconstruction residual thereupon determines its identity or expression. We refer readers to  for the exact problem set-up and implementation details. Table 2 collects the computed recognition rates. Although RPCAG and FR- PCAG are superior than PCP as expected, PCPS performs distinctly better than all others.
We initialized the weights randomly between -0.5 to 0.5. The uni-polar sigmoid activation function shown in equation 6 is selected comparing the accuracy rate of different function in . The actual output of the neurons in hidden layer and output layer is calculated by activation function using forward propagation. The error gradient is calculated using the actual output and the desired output. The error is propagated backwards in the network simultaneously calculating weight correction. Finally, all the weights are updated and this is repeated for each training sample in all epochs. The error gradient is minimized in each iteration using fmincg function of MATLAB.
In face recognition, the images of human face is dimensionality high and requires a longer time to perform the classification task. This problem can be solved by reducing the dimension of the face images thus PrincipalComponentAnalysis (PCA) is being recommended because it can significantly reduces the dimension of human faces. The PCA is being introduced by M.A Turk and Alex P.Pentland in 1991 and they have developed a near real time Eigenface system for face recognition by computing the Euclidean distance. While reducing the dimension of facial images, the deviation of the facial images are being retain at the same time. Basically, the PCA method is based on information theory approach that fractionate the facial images into a set of characteristic feature images as known as Eigenfaces and these Eigenfaces are considered as the principal components of the initial training set of images. There are a few steps when performing face recognition, firstly a new input image is being projected into the “face space”, a subspace that spanned by the eigenfaces. Then, the face are being classified by comparing its location in the face space with the locations of known person.
We discuss a class of principalcomponent analyzers defined using generic functions which con- tain tuning parameters. For example if we adopt a log-sigmoidal function as a generic function, the tuning parameters are the inverse temperature and saturation value parameters, as will be discussed in detail. In general the tuning parameter set makes a delicate trade between loss of information and degree of insensitivity to outliers. The main objective in the present paper is to provide a reasonable selection of tuning parameters of principalcomponent analyzers. The basic idea is to craft a loss function that reflects as appropriate trade off between loss of information and robustness to outliers. We introduce K-fold cross validation for estimating the expected loss based on a given data set. As a result we build a method of data-adaptive selection of tuning parameters. In a simulation study, we examine the performance of the adaptive selection under three types of outlier distributions H displaying deterministic, structural and distributional contaminations based on (2). The three types of outliers are simulated in a numerical experiment, and we test the performance in a few cases of principalcomponent analyzers. We provide an S implementation of the basic robust PCA at http://home.hiroshima-u.ac.jp/oxbow/RobustPCA/.
After extraction of the needed features, we apply the PCA to the same image set in the database, this gives us some weight descriptors for each image after the eigenface has been computed and thus gives us the possibility of adding the total weight score we got for each subject from their fingerprints to the new score computed based on their weight from the eigenface. For any probe image we are recognizing, we also make it to go through the above steps, such that the important features are also extracted and scored and the eigenface computed. We finally add the two weight scores i.e. from the eigenface computation and the features extracted and then compare the score with the aggregate scores in the database, if it matches the score of any subject in the database, we recognize it as known. The steps are detailed as below: 1) Face database formation phase: Acquisition and pre- processing / normalization of face images are done here, then the images are stored in the database. Training is performed on the images in this database and their corresponding eigenfaces and eigenvalues created. Our system is designed to operate on 128 x 128 images in the database, to perform image size conversions and enhancements on face images; we have pre-processing steps for normalization which rescale all images to 128 x 128. Here, we also perform histogram equalization and background removal to improve face recognition performance. For each face acquired, we have two entries in the database such that one is for the image itself while the other is the weight vectors computed after training is done; this weight vector is used to compute the ultimate weight for each image.
Face recognition from still images and video sequence has been an active research area due to both its scientific challenges and wide range of potential applications such as biometric identity authentication, human-computer interaction, and video surveillance. Within the past two decades, numerous face recognition algorithms have been proposed as reviewed in the literature survey. Even though we human beings can detect and identify faces in a cluttered scene with little effort, building an automated system that accomplishes such objective is very challenging. The challenges mainly come from the large variations in the visual stimulus due to illumination conditions, viewing directions, facial expressions, aging, and disguises such as facial hair, glasses, or cosmetics . Face Recognition focuses on recognizing the identity of a person from a database of known individuals. Face Recognition will find countless unobtrusive applications such as airport security and access control, building surveillance and monitoring Human-Computer Intelligent interaction and perceptual interfaces and Smart Environments at home, office and cars .
A study has been conducted by Safaei and Wu (2015) in order to evaluate 3D hand mo- tion with SoftKinetic. Hidden Markov Models were used to conduct the classification. The proposed method uses a framework to capture RGB and depth data from a SoftKinetic camera and a UV map 5 generated to merge the RGB image with the depth image to create a 3D model. Chaczko and Alenazy (2016) discuss modelling and simulation of gesture recognition technologies. The initial plan was to use a Kinect camera. However, this plan was changed to use the SoftKinetic DepthSense 325 camera due to ease of use and compatibility with the IISU middleware. The goal was the development of a web application that can monitor users, who are making gestures in front of the camera(s), and that can provide controls for other possible applications. However, it is not clear what kind of gestures the web application is able to recognise nor how the process of classification is made.
This algorithm is limited by the requirement of com- plex k-space data, which is not always saved in clinical studies. Thus, MRI operators must be vigilant to identify corrupted datasets and save the raw data for postprocess- ing, if this is not automatically done by default. Future studies will investigate the application of this algorithm to magnitude images, more commonly saved, by Fourier transforming to produce a pseudo-k-space (1). In the pseudo-k-space, the spikes produce a strong component at k¼0, making it more difficult to differentiate the genu- ine central peak from RF spike contributions with good sensitivity. In general, multi-frame datasets are preferred for this decomposition, because the algorithm can accu- rately identify the center of k-space as consistent between frames and not misclassify it as an RF spike. Similarly, the power of the decomposition is somewhat reduced if the spikes have consistent structure between frames. Many common clinical and research imaging methods rely on a large number of images (e.g., functional MRI, DTI, arterial spin labeling, cardiac function imaging), and the presented RPCA method for correcting k-space spike artifacts is expected to have widespread use.
The training file will contain m eigen weights for each of the m images. This is written in a training file with each of the image name followed by its id and its eigen weights.The training file is train.txt.The identification image also will have m eigen weights for recognition which is written in the file out.txt.
An attempt is made to design a universal robust face recognition methodology which is invariant to changes in pose, illumination and expressions as well. To address it, we have proposed a design using Kernel Entropy ComponentAnalysis based on Gaussian non- additive entropy measure. We have used Gabor Wavelet Transformation for achieving illumination invariance with some degree of expression and pose invariance, and Discrete Cosine Transform for significant robustness towards illumination changes in the face. The methodology not only yields high success rate due to discriminatory power of the used entropy measure and the Inner Product Classifier, but is also computationally efficient as a common solution to pose, illumination and expression variations. Using NNS, each fiducial point in the test feature vector is compared with the fiducial points in every training feature vector. The maximum number of matches between the test and the training feature vectors leads to a possible match.
Principle componentanalysis (PCA) is used for recognition which takes the sample image and encodes it in the same way and compares it with the set of encoded images. There can be many Eigen vectors for a covariance matrix but very few of them are principle one‘s. Each Eigen vector can be used for finding different amount of variations among the face image. The goal of PCA is to reduce the dimensionality of the data while retaining as much as possible of the variation present in the original dataset. PCA allows us to compute a linear transformation that maps data from high dimensional space to low dimensional sub-space.
Abstract—In this paper we give a comparative analysis of performance of feed forward neural network and elman neural network based face recognition. We use different inner epoch for different input pattern according to their difficulty of recognition. We run our system for different number of training patterns and test the system’s performance in terms of recognition rate and training time. We run our algorithm for face recognition application usingPrincipalComponentAnalysis and both neural network. PCA is used for feature extraction and the neural network is used as a classifier to identify the faces. We use the ORL database for all the experiments. Here 150 face images from the database are taken and some performance metrics such as recognition rate and total training time are calculated. We use two way cross validation approach while calculating recognition rate and total training time . In two way cross validation, we interchange training set into test set and test set into training set. Feed forward neural network has better performance in terms of recognition rate and total training time as compare to elman neural network.
of the sample trimmed PC space towards the population counterpart. The ro- bustness of the method is studied by showing qualitative robustness, computing the breakdown point, and deriving the influence functions, which turn out to be bounded for bad leverage points. Good leverage points still may have an un- bounded influence. Furthermore, asymptotic efficiencies at the normal model are derived, while finite sample efficiencies of the estimators are obtained by means of a simulation study. It is shown that, by selecting an appropriate trimming proportion α, both a high breakdown point and a high efficiency are attainable. A distinct feature of the proposed method compared to other approaches for robust PCA is that it directly aims at finding the best fitting affine subspace. The population version, which we presented in Section 2 and of which we showed existence in Section 3, has a clear geometric interpretation, also at non-elliptical distributions. If one would use, for example, the space spanned by the first d eigenvectors of a robust estimate of the covariance matrix as best fitting sub- space, then it is not clear whether the corresponding population quantity has any optimality property, unless at elliptically symmetric distributions. When the aim of the robustprincipalcomponentanalysis is to perform dimension re- duction and to find an optimal subspace of a certain dimension, then trimmed PCA is a natural candidate. A plot of the values of the trimmed variation as a function of d can be used to select the dimension of the subspace. If such a plot indicates that not much further reduction in trimmed variation can be gained by increasing d to d + 1, the corresponding dimension can be selected.
the human fillings and emotions. Outward appearance is movements or positions of the muscles underneath the skin of the face. Though nothing is said verbally, there is much to be understood about the messages we send and receive through the use of nonverbal communication, such our expressions. Such as communication is two types Verbal and Non-Verbal in this two kind of correspondence (contact) outward appearance is sort of non-verbal correspondence yet it expects critical part. The feelings expressed on a person's face; a sad expression, a look of triumph, an angry face .This paper incorporates presentation of facial feeling acknowledgment framework, Application, relative investigation of well known face look acknowledgment methods and periods of programmed outward appearance acknowledgment framework. Facial expression plays a principal role in human interaction and communication since it contains critical and necessary information regarding emotion. This paper goal is to present needs and utilizations of outward appearance acknowledgment.
Over the last ten years or so, face recognition has become a popular area of research in computer vision and one of the most successful applications of image analysis and understanding. Because of the nature of the problem, not only computer science researchers are interested in it, but neuroscientists and psychologists also. It is the general opinion that advances in computer vision research will provide useful insights to neuroscientists and psychologists into how human brain works, and vice versa .The goal is to implement the system (model) for a particular face and distinguish it from a large number of stored faces with some real-time variations as well. It gives us efficient way to find the lower dimensional space. Further this algorithm can be extended to recognize the gender of a person or to interpret the facial expression of a person. Recognition could be carried out under widely varying conditions like frontal view, a 45° view, scaled frontal view, subjects with spectacles etc are tried, while the training data set covers limited views. The algorithm models the real-time varying lighting conditions as well. But this is out of scope of the current implementation .