Search

Wearable-based approach and vision-based approach are two of the most common approaches in human activity recognition. However, the concern of privacy issues may limit the application of the vision-based approach. Besides, some individuals are reluctant to wear sensor devices. Hence, smartphone-based human physical activity recognition is a popular alternative. In this paper, we propose a deep analysis to interpret and predict accelerometer data captured using a smartphone for activity recognition. The proposed deep model is able to extract deep features from both spatial and temporal domains of the inertial data. The recognition accuracy of the proposed model is assessed using UCI and WISDM accelerometer data. Empirical results exhibit a promising performance, attaining accuracy score of 90% in UCI and 87% in WISDM database.

The Deep learning analysis has been extensively carried out in the context of object/ pattern recognition due to its excellence in feature extraction and classification. However, the superior performance just can be guaranteed with the availability of huge amounts of training data and also high-specification data processing unit to process the data deeper at high speeds. Hence, another alternative is by applying transfer learning. In transfer learning, a neural network model is first trained on a data similar to the targeted data. With that, the knowledge such as features, weights etc. could be leveraged from the trained model to train the new model. In this project, a vision-based human action recognition via a transfer learning is conducted. Specifically, in the proposed approach, the earlier layers of a pre-trained AlexNet is preserved since those extracted low-level features are characterizing generic features which are common to most data. However, the pre-train network is fine-tuned based on our interested data, that is human action data. Since AlexNet requires input data of size 227*227*3, the frames of each video are processed into 3 different templates. The three computed templates are: (1) Motion History Image carrying spatio-temporal information, (2) Binary Motion Energy Image incorporating motion region information and (3) optical flow template holding accumulative motion speed information. The proposed approach is validated on two publicly available databases, which are Weizmann database and KTH database. From the empirical results, a promising performance is obtained with about 90% accuracy from the databases.

The development of automated video surveillance has grown dramatically due to the increased concern with public safety and security. An automated surveillance with reliable human activity analysis is essential. In this paper, a localized spatio-temporal representation, alongside Motion History Image (MHI), Motion Energy Image (MEI) and Binarized Statistical Image Features (BSIF), is proposed for human action recognition. In this work, the information of timestamp and ratio of colors are extracted from the silhouette of MHI template. This information is then utilized to derive a temporal representation for encoding movement dynamics. This temporal representation preserves transient information of actions. Subsequently, local descriptors are computed from MHI and MEI temporal templates via BSIF. The computed localized temporal representation is classified by using a linear SVM. The proposed system offers promising performance in human action recognition with about 90% accuracy.

In this paper, we propose a simple and robust local descriptor operator, called the extended local graph structure (ELGS). The original local graph structure (LGS) performs very well in many domains, for instance face recognition, face spoofing detection and others. However, LGS has a few demerits such as LGS is not robust to the noise present in the image and LGS takes into considerations the horizontal graph and ignores the vertical graph which causes a loss in the spatial information. Therefore, we extend the idea of LGS by encoding the pattern into two directions. This is means that we take into consideration the vertical graph along with the horizontal graph and then concatenate the two computed histograms features to form a global descriptor. Experimental results on the UIUC and XU High Resolution texture databases show a promising performance.

This paper devises a filter bank approach to extract the local structure knowledge of a face image by computing its probability distribution function from the filter responses. Independent Component Analysis (ICA) filters are embraced in this work. Considering the limitation of ICA filter learning in handling the image conditions with uncontrolled facial expressions, illuminations, aging, etc., we proposed an independent statistical descriptor, coined ISD, in this paper with the aim to handle the rampant images. ISD intensifies ICA response invariance through hashing the filter responses by encoding the relation between each response element and its neighbour and then block-wise histogramming the output. In addition, an overlapping average pooling is executed to regulate the histogram features, prior to whitening PCA compression. The good performance of ISD descriptors has been extensively corroborated in the empirical results on face recognition.

In this paper, we present a novel and effective feature extraction technique for face recognition. The proposed technique incorporates a kernel trick with Graph Embedding and the Fisher's criterion which we call it as Kernel Discriminant Embedding (KDE). The proposed technique projects the original face samples onto a low dimensional subspace such that the within-class face samples are minimized and the between-class face samples are maximized based on Fisher's criterion. The implementation of kernel trick and Graph Embedding criterion on the proposed technique reveals the underlying structure of data. Our experimental results on face recognition using ORL, FRGC and FERET databases validate the effectiveness of KDE for face feature extraction.

Graph-based subspace learning is a class of dimensionality reduction technique in face recognition. The technique reveals the local manifold structure of face data that hidden in the image space via a linear projection. However, the real world face data may be too complex to measure due to both external imaging noises and the intra-class variations of the face images. Hence, features which are extracted by the graph-based technique could be noisy. An appropriate weight should be imposed to the data features for better data discrimination. In this paper, a piecewise weighting function, known as Eigenvector Weighting Function (EWF), is proposed and implemented in two graph based subspace learning techniques, namely Locality Preserving Projection and Neighbourhood Preserving Embedding. Specifically, the computed projection subspace of the learning approach is decomposed into three partitions: a subspace due to intra-class variations, an intrinsic face subspace, and a subspace which is attributed to imaging noises. Projected data features are weighted differently in these subspaces to emphasize the intrinsic face subspace while penalizing the other two subspaces. Experiments on FERET and FRGC databases are conducted to show the promising performance of the proposed technique.

In this paper, we present an effective technique on discriminative feature extraction for face recognition. The proposed technique incorporates Graph Embedding and the Fisher's criterion where we call it as Neighbourhood Preserving Discriminant Embedding (NPDE). Utilizing the Graph Embedding criterion, the underlying nonlinear face data structure is revealed as representative and discriminative features for analysis. We employ Neighbourhood Preserving Embedding (NPE) for the purpose. NPE takes into account the restriction that neighbouring points in the high-dimensional space must remain within the same neighbourhood in the low dimension space and be located in a similar relative spatial situation (without changing the local structure of the nearest neighbours of each data point). Furthermore, by taking the advantage of the Fisher's criterion, the discriminating power of NPDE is further boosted. Based on this intuition, NPDE obtains better discriminative capability and experimentally verified in ORL, PIE and FRGC.

A mobile application of in-air signature is proposed by implementing the algorithm of Fast Dynamic Time Warping (FastDTW). With the linear time- and memory-complexity of FastDTW, the authentication process can be done on a real-time mode. The programmed mobile application was tested on 50 volunteers, and 5 fraudsters who contributing a total of 500 genuine in-air signature samples, and 1250 forged in-air signature samples. Average error rate (AER) of 0.098 is reported based on the private database. This promising result shown the compatibility of FastDTW in mobile application. The experienced users' feedbacks also shown the positive acceptability of this proposed mobile application.

Semi-supervised learning approach is a fusion approach of supervised and unsupervised learning. Semi-supervised approach performs data learning from a limited number of available labelled training images along with a large pool of unlabelled data. Semi-supervised discriminant analysis (SDA) is one of the popular semi-supervised techniques. However, there is room for improvement. SDA resides in the illumination and local change of the face features. Hence, it is hardly to guarantee its performance if there are illumination and local changes on the images. This paper presents an improved version of SDA, termed as Semi-Supervised Discriminant Local Analysis (SDLA). In this proposed technique, a local descriptor is amalgamated with SDA. Hence, SDLA could possess the capabilities of both the local descriptor and SDA, in such a way that SDLA utilizes limited number of labelled training data and huge pool of unlabelled data to optimally capture local discriminant features of face data. The empirical results demonstrate that SDLA shows promising performance in both normal and makeup face authentication.

The Internet has always been growing with all the contents and information added by different types of users. Without proper storage and indexing, these contents can easily be lost in the sea of information housed by the Internet. Hence, an automated program, known as the web crawler is used to index all the contents added to the Internet. With proper configurations and settings, a web crawler can be used for other purposes besides web indexing, which include downloading files from the web. Millions or billions of files are uploaded on the Internet and for most of the sites which host these files, there are no direct indication of whether the file is safe and free of malicious codes. Therefore, this paper aims to provide a construction of a web crawler which crawls all the pages in a given website domain, and download all the possible downloadable files linked to those pages, for the purpose of file safety inspection.

Supervised learning techniques are preferable in face recognition for their pleasant data discriminating capability. However, their performance just can be assured if and only if there are sufficient labelled training images available. Practically, it always happens that only a small number of labelled training images available due to costly and time consuming labelling process. On the other hand, a large pool of unlabeled data could be easily obtained through public databases like Google or Flickr. Hence, semi-supervised learning is an alternative direction in face recognition. Semi-supervised techniques utilize limited labelled training images and huge amount of unlabeled training data for data learning. This paper presents a new semi-supervised technique, namely Semi-supervised Generic Descriptor (SSGD). SSGD uses labelled training images to compute the null space of class scatter vector and generate class generic descriptors to represent each class. Besides that, unlabelled training images are exploited to obtain more information about face data structure. The empirical results demonstrate that SSGD shows relatively promising performance in face verification.

Discriminant Common Vectors (DCV) is proposed to solve small sample size problem. Face recognition encounters this dilemma where number of training samples is always smaller than the data dimension. In literature, it is shown that DCV is efficient in face recognition. In this paper, DCV is enhanced for further boosting its discriminating power. This modified version is namely Discriminative Discriminant Common Vectors (DDCV). In this technique, a local Laplacian matrix of face data is computed. This matrix is used to derive a regularization model for computing discriminative class common vectors. Experimental results demonstrate that DDCV illustrates its effectiveness on face verification, especially on facial images with significant intra class variations.

Semi-supervised discriminant analysis (SDA) is a popular semi-supervise learning technique for limited labelled training sample problem in face recognition. However, SDA resides in the illumination variations and noise of the face features. Hence, SDA exposes the illumination variations and noise when constructing the optimal projection. It could affect the projection, leading to poor performance. In this paper, an enhanced SDA, namely Wavelet SDA, is proposed. This proposed technique is to resolve the problem of intra-class variations due to illumination variations and noise on image data. The robustness of the proposed technique is evaluated using three well-known face databases, i.e. ORL, FERET and FRGC. Empirical results validated the good effects of wavelet transform on SDA, leading to better recognition performance.

Face spoofing attack is one of the recent security traits that face recognition systems are proven to be vulnerable to. The spoofing occurs when an attacker bypass the authentication scheme by presenting a copy of the face image for a valid user. Therefore, it's very easy to perform face recognition spoofing attack with compare to other biometrics. This paper, addresses the problem of detecting imposter face image from live image. In practically, we address this problem from texture analysis point of view because the printed face usually has less quality defect that can be observed by extracting texture features. We adopt Local graph structure LGS to extract the features. Moreover, LGS is based on applying a dominant graph into the input image and it's proved to be a powerful texture operator. Finally, extensive experimental analysis on NUAA showed an encouraging performance. Index Terms - Local Graph Structure, image processing, pattern recognition, face recognition, face spoofing.

Face recognitions systems suffer from the problem associated with illumination variation. Therefore, there's a need to address this problem. In this paper, we present a novel algorithm for illumination normalization call Local Trapezoid Feature LTF. The features are derived from the trapezoid rule and the experiments results on extended Yale face database demonstrated the effectiveness and the superiority of the algorithm. Furthermore, our algorithm doesn't require dimensionality reduction or feature extraction.

In this paper, we proposed and evaluated a new algorithm (EdgePass) to extract the edges of pass-images for a graphical password system. The EdgePass algorithm is based on an estimated technique for calculating the area in a plane bounded by a graph. This proposed scheme is the first stage of our graphical password system. The system framework is based on Edge detection, Image fusion and high spatial frequency component extraction. The idea of the proposed algorithm is to find whether the process pixel I(x, y) is dark or light. If it is light then the corresponding target pixel I(x, y) is an edge, else it is a background. The experiment results demonstrated the effectiveness of the proposed method. The new method can be stabilized more quickly.

Locally Linear Embedding (LLE) is a popular dimension reduction technique due to its nonlinearity property. However, LLE is restricted to its unsupervised nature and “out-of-sample problem” in which suboptimal to face recognition problem. Hence, we propose a supervised and linear approximation of LLE, known as Neighborhood Preserving Discriminant Embedding (NPDE). Using the class information, NPDE finds an optimal projection so that the ratio of the within-neighborhood scatter and the between-neighborhood scatter is minimized. NPDE signifies the local neighboring geometry that corresponding to the nonlinear underlying data structure in the image space. Based on this intuition, NPDE shows better discriminative capability in face recognition.

Neighborhood Preserving Embedding (NPE) is an unsupervised linear dimensionality reduction technique which attempts to solve the ldquoout of samplerdquo problem in Locally Linear Embedding (LLE). This is done by introducing a linear transform matrix into LLE, and hence NPE can be perceived as a linear approximation to LLE. In this paper, we modify the original NPE for face recognition by embedding prior class information in the process of neighborhood selection. Intuitively, neighboring points are kept intact if they have the same class label, while avoid points of other classes from entering the neighborhood. We proved experimentally in three face databases, ie. ORL, PIE and FRGC, and with comparisons with other linear and non-linear feature extractors, the intuition underlying the inclusion of class information in NPE works out very advantageously for achieving high recognition performance.

Face images are often very high-dimensional and complex. However, the actual underlying structure can be characterized by a small number of features. Hence, locally linear embedding (LLE) is proposed as a nonlinear dimension reduction technique to deal this problem. LLE learns the intrinsic manifold embedded in the high dimensional ambient space by minimizing the global reconstruction error of the neighbourhood in the data set. LLE is popular in analyzing face images with different poses, illuminations or facial expressions for one subject class. It is developed based on the assumption that data that is distributed on a single manifold is having the same class label; hence the process of neighborhood selection is non class-specific. However, this is inappropriate to face recognition as face recognition learns in multiple manifolds where each representing data on one specific class. Here, we modify the original LLE by embedding prior class information in the process of neighborhood selection. Experimental results demonstrate that our technique consistently outperforms the original LLE in ORL, PIE and FRGC databases.

This paper presents a face recognition technique with two techniques: random projection (RP) and robust linear discriminant analysis model (RDM). RDM is an enhanced version of fisher's linear discriminant with energy-adaptive regularization criteria. It is able to yield better discrimination performance. Same as Fisher's Linear Discriminant, it also faces the singularity problem of within-class scatter. Thus, a dimensionality reduction technique, such as principal component analysis (PCA), is needed to deal with this problem. In this paper, RP is used as an alternative to PCA in RDM in the application of face recognition. Unlike PCA, RP is training data independent and the random subspace computation is relatively simple. The experimental results illustrate that the proposed algorithm is able to attain better recognition performance (error rate is approximately 5% lower) compared to Fisherfaces.

Feature extraction is a data analysis technique devoted to removing redundancy and extracting the most discriminative information. In face recognition, feature extractors are normally plagued with small sample size problems, in which the total number of training images is much smaller than the image dimensionality. Recently, an optimized facial feature extractor, maximum marginal criterion (MMC), was proposed. MMC computes an optimized projection by solving the generalized eigenvalue problem in a standard form that is free from inverse matrix operation, and thus it does not suffer from the small sample size problem. However, MMC is essentially a linear projection technique that relies on facial image pixel intensity to compute within- and between-class scatters. The nonlinear nature of faces restricts the discrimination of MMC. Hence, we propose an improved MMC, namely maximum neighborhood margin criterion (MNMC). Unlike MMC, which preserves global geometric structures that do not perfectly describe the underlying face manifold, MNMC seeks a projection that preserves local geometric structures via neighborhood preservation. This objective function leads to the enhancement of classification capability, and this is testified by experimental results. MNMC shows its performance superiority compared to MMC, especially in pose, illumination, and expression (PIE) and face recognition grand challenge (FRGC) databases.