Figure 3 - uploaded by Stan Z Li
Content may be subject to copyright.
Source publication
Recently, we developed NIR based face recognition for highly accurate face recognition under illumination varia- tions (10). In this paper, we present a part-based method for improving its robustness with respect to pose variations. An NIR face is decomposed into parts. A part classifier is built for each part, using the most discriminative LBP his...
Context in source publication
Context 1
... α t ∈ R are the combining weights. We can consider the real-valued number T t =1 α t h t ( x ) as the score, and make a decision by comparing the score with a thresh- old. An AdaBoost learning procedure is aimed to derive α t and h t ( x ) so that an upper error bound is minimized [5]. The procedure keeps a distribution w t = ( w t, 1 , . . . , w t,N ) for the training examples. The distribution is updated after each learning iteration t . The AdaBoost procedure adjusts the distribution in such a way that more difficult examples will receive higher weights. While an AdaBoost procedure essentially learns a two- class classifier, we convert the multi-class problem into a two-class one using the idea of intra- and extra-class difference [12]. However, here the difference data are derived from the SLBPH features rather than from the images. A difference is taken between two SLBPH feature sets, which is intra-class if the two face images are of the same person, or extra-class if not. to construct weak classifier for the above AdaBoost learning, we generate samples using histogram bin difference as dissimilarity measure between SLBPH features of two faces. To achieve a high accurate face recognition system, more effort should be made to overcome some scenario changes, such as pose and expression variations. In this paper, we propose a part-based solution to this aim. Firstly, according to a statistical analysis, we decompose the NIR face image into several parts, such as eyes, noses, and mouth. Then, AdaBoost learning is applied to each part to build a part- based classifier, and finally the output of each part classifier is fused to give the final score. Human face consists of several different parts, such as eyes, nose, and mouth. Earlier researches show that each facial part has a different contribution in face recognition [8, 2]. In this paper, we propose a statistical learning method to analyze the discriminant power of each facial part. Using SLBPH as feature representation, AdaBoost [4] can be further applied to learn the most discriminative features for face recognition. Since each learned SLBPH feature covers a subregion of the whole face, areas covered by more SLBPH features have more importance in face recognition. We select the first 100 SLBPH features learned by AdaBoost to demonstrate how facial parts contributes in face recognition. Given each subwindow of SLBPH feature with equal gray intensity, we further overlap them all in the same face according to their locations. Then the brightest parts indicate the most discriminative power for face recognition. The result is shown in Fig. 2. According to the above analysis, we decompose the whole face into several facial parts. When a new image comes into our system, we first detect the face, and normal- ize it to a fixed size according to the eye coordinates. Then, we decompose the face into several different parts. The pro- cess is shown in Fig. 3. When we decompose the face into parts, the size of subwindows are restricted by the size of each part. The maximum size of the subwindows is the size of the biggest part. It can be seen that, in whole face, there are some big subwindows that are larger than any decomposed part. These subwindows are definitely lost if we only consider the subwindows in each part. We can also find out that some subwindows covers more than one part. If we only consider the subwindows of each part, these subwindows will also be lost. To make it up, we consider the whole face as one big facial part. In this system, a strong AdaBoost classifier is learned for each facial part, so that each part classifier gives out a score. The final result is reached by fusions of all part classifiers. Before fusion, we apply Z score normalization to each output of part classifiers, so that they have zero mean and unit standard variance individually. Kittler et al. [9] have developed a theoretical framework for consolidating the evidence obtained from multiple classifiers using schemes like the sum rule, product rule, max rule, min rule, median rule and majority voting. In this paper, we use sum rule and max rule to do score fusion and compare the performances. The max rule and sum rule are used for score fusion in this paper. The max rule approximates the mean of the posterior probabilities by the maximum value, so the final score is given ...
Similar publications
The interest in variable selection for clustering has increased recently due
to the growing need in clustering high-dimensional data. Variable selection
allows in particular to ease both the clustering and the interpretation of the
results. Existing approaches have demonstrated the efficiency of variable
selection for clustering but turn out to be...
Metabolomic studies with a time-series design are widely used for discovery and validation of biomarkers. In such studies, changes of metabolic profiles over time under different conditions (e.g., control and intervention) are compared, and metabolites responding differently between the conditions are identified as putative biomarkers. To incorpora...
Feature selection is an important component of text categorization. This technique can both increase a classifier's computation speed, and reduce the overfitting problem. Several feature selection methods, such as information gain and mutual information, have been widely used. Although they greatly improve the classifier's performance, they have a...
The properties of road networks vary considerably from one geographical environment to another. The networks pertaining in a satellite image can therefore be used to classify and retrieve such environments. In this work, we propose to classy geographical environment using geometrical and topological features computed from the road networks. The lim...
The CVLT-II provides standardized scores for each of the List A five learning trials, so that the clinician can compare the patient's raw trials 1-5 scores with standardized ones. However, frequently, a patient's raw scores fluctuate making a proper interpretation difficult. The CVLT-II does not offer any other methods for classifying a patient's l...
Citations
... Researchers from different corners of the world have attempted to recognize facial expressions using ML-based [9][10][11] and DL-based [12][13][14][15] approaches, where they have considered IR facial images as the input to their models. However, it has been noticed that the performance of DL-based approaches is better than ML-based approaches due to the automatic extraction of powerful features from inputs by DL models [12]. ...
... In the literature, part-based approaches have been employed, mostly for 2D images, in which separate models are trained for local regions such as the eyes, nose, mouth and chin, along with the entire face [34], [35], [36], [37]. Defining local regions on high resolution 3D facial meshes is challenging. ...
Face recognition is a widely accepted biometric identifier, as the face contains a lot of information about the identity of a person. The goal of this study is to match the 3D face of an individual to a set of demographic properties (sex, age, BMI, and genomic background) that are extracted from unidentified genetic material. We introduce a triplet loss metric learner that compresses facial shape into a lower dimensional embedding while preserving information about the property of interest. The metric learner is trained for multiple facial segments to allow a global-to-local part-based analysis of the face. To learn directly from 3D mesh data, spiral convolutions are used along with a novel mesh-sampling scheme, which retains uniformly sampled points at different resolutions. The capacity of the model for establishing identity from facial shape against a list of probe demographics is evaluated by enrolling the embeddings for all properties into a support vector machine classifier or regressor and then combining them using a naive Bayes score fuser. Results obtained by a 10-fold cross-validation for biometric verification and identification show that part-based learning significantly improves the systems performance for both encoding with our geometric metric learner or with principal component analysis.
... Ke Pan et al. [25] developed a part based facial recognition system through NIR (Near Infrared Images) for extremely accurate and efficient facial recognition system under light variations. This approach is also for improving robustness with respect to pose variations. ...
... Composite sketches are commonly used in the criminal investigation to assist in the recognition of suspect involved in criminal activities. There are two facial sketch to mugshot matching algorithms being used in recognition systems: a holistic method [15] and a component-based method [25]. Most of the previous studies focused on the component-based method but none of them uses the Weighted Component-Based Approach. ...
Facial recognition is a popular biometric technique to recognize an individual by comparing the facial features of a given photograph or a sketch to the digitally stored photographs. One of the important applications of facial recognition is to determine the identity of criminals through their hand-drawn or composite sketches. Despite of the development made in sketch-based facial recognition, available approaches are facing various challenges. The component-based approach (CBA) measures the similarity between each facial component of a sketch and a mugshot photograph. The major challenge in this approach is to determine which facial components are crucial in the identification process. Certain facial components provide better recognition clue than others while matching with mugshot photographs and considerably accurate identification results could be achieved to incorporate such crucial components in the recognition process. In this article, we propose a novel methodology which is based on computable weights to find the most discriminative facial components by using the Weighted Component-Based Approach (WCBA). The weight vector is used during the similarity score measurement to enhance the accuracy and performance of the facial recognition system. Experimental results on matching 50 facial images from 1193 subjects of Multiple Encounter Dataset II (MEDS-II) and 85 facial images from CHUK face sketch database (CUFS) show that the proposed method achieves promising performance (accuracies of 58.33% and 88.23%, respectively) as compared to other leading facial recognition techniques (accuracies of 52% and 80%). We believe our prototype approach will be of great value to law enforcement agencies in the apprehension of culprits in a timely fashion.
... Data Augmentation/ Recovery [15], [20], [21], [24], [26], [31], [35], [51], [54], [62], [64], [73], [78], [89], [90], [107], [112], [136], [140], [147], [183] Feature Extraction [1], [2], [11], [23], [27], [32], [33], [48], [49], [53], [60], [62], [65], [74], [80], [88], [93], [108], [109], [124], [125], [128], [134], [143], [148], [153], [162], [163], [169], [180], [187], [191], [193] Feature Comparison [4], [19], [55], [57], [96], [98], [101], [133], [150]- [152] Fusion Strategy [12], [46], [50], [103], [111], [122], [135], [185], [187] ...
... Application-oriented Purpose Publication OFD Occluded Face Detection [14], [18], [36], [61], [61], [91], [106], [120], [146], [155], [159], [165], [165]- [167], [174], [192] ORFE Patch based engineered features [1], [2], [12], [23], [57], [65], [88], [111], [122], [180], [181], [193] Learning based features [11], [27], [31]- [33], [53], [65], [73], [74], [90], [93], [96], [107], [109], [124], [125], [133], [134], [136], [148], [153], [162], [163], [169], [187] OAFR Occlusion Detection [18], [156] Occlusion Discard Face Recognition [4], [19], [98], [101], [103], [108], [128], [143] Partial Face Detection NA Partial Face Recognition [46], [48]- [50], [55], [80], [150], [151] ORFR Occlusion Recovery [15]- [17], [20], [21], [29], [51], [54], [59], [64], [77], [83], [114], [139], [140], [157], [161] Occlusion Recovery Face Recognition [24], [26], [31], [35], [60], [62], [73], [78], [89], [112], [147], [183], [185], [191] ...
The limited capacity to recognize faces under occlusions is a long-standing problem that presents a unique challenge for face recognition systems and even for humans. The problem regarding occlusion is less covered by research when compared to other challenges such as pose variation, different expressions, etc. Nevertheless, occluded face recognition is imperative to exploit the full potential of face recognition for real-world applications. In this paper, we restrict the scope to occluded face recognition. First, we explore what the occlusion problem is and what inherent difficulties can arise. As a part of this review, we introduce face detection under occlusion, a preliminary step in face recognition. Second, we present how existing face recognition methods cope with the occlusion problem and classify them into three categories, which are 1) occlusion robust feature extraction approaches, 2) occlusion aware face recognition approaches, and 3) occlusion recovery based face recognition approaches. Furthermore, we analyze the motivations, innovations, pros and cons, and the performance of representative approaches for comparison. Finally, future challenges and method trends of occluded face recognition are thoroughly discussed.
... Although many studies [3,4] have considered these advantages with respect to NIR face recognition, few studies have applied deep learning to NIR face recognition. Some studies have used deep learning for cross-modality face recognition between RGB, NIR, and thermal. ...
... Among these facial variations, wearing eyeglasses makes recognition difficult for NIR face images. If people wear glasses, the active NIR light is reflected in the glasses and the reflected light covers the eyes, as shown in Figure 2. According to the literature [4], the area around the eyes is the most discriminative area for differentiating between faces. Thus, traditional face recognition methods which focus on the eye regions are hindered by artifacts covering the eye region. ...
Face recognition using a near-infrared (NIR) sensor is widely applied to practical applications such as mobile unlocking or access control. However, unlike RGB sensors, few deep learning approaches have studied NIR face recognition. We conducted comparative experiments for the application of deep learning to NIR face recognition. To accomplish this, we gathered five public databases and trained two deep learning architectures. In our experiments, we found that simple architecture could have a competitive performance on the NIR face databases that are mostly composed of frontal face images. Furthermore, we propose a data augmentation method to train the architectures to improve recognition of users who wear glasses. With this augmented training set, the recognition rate for users who wear glasses increased by up to 16%. This result implies that the recognition of those who wear glasses can be overcome using this simple method without constructing an additional training set. Furthermore, the model that uses augmented data has symmetry with those trained with real glasses-wearing data regarding the recognition of people who wear glasses.
... Using the entire face as a global feature is weak in occlusion, such as wearing glasses or sunglasses. Furthermore, there are methods to extract features by dividing the face locally [8,9]. In recent years, deep-learning based methods that train a large amount of data and improve recognition accuracy has become an issue. ...
... The local feature is robust to occlusion comparing to the global feature due to regional sampling. When extracting features, Local Binary Pattern (LBP) [10] and the Histogram of Oriented Gradient (HOG) [13] are mainly used for local descriptors [8,14,15]. The LBP feature assigns a feature value to eight neighbors of a pixel, using the value of this center pixel as a threshold. ...
... However, the computational cost of each algorithms is expensive and the required alignment step limits its practical applications. Besides, region-based models [3], [9], [23], [24], [25], [32], [33] also offered a solution for partial face recognition. They only required face sub-regions as input, such as eye [32], nose [32], half (left or right portion) of the face [9], or the periocular region [26]. ...
Biometric recognition on partial captured targets is challenging, where only several partial observations of objects are available for matching. In this area, deep learning based methods are widely applied to match these partial captured objects caused by occlusions, variations of postures or just partial out of view in person re-identification and partial face recognition. However, most current methods are not able to identify an individual in case that some parts of the object are not obtainable, while the rest are specialized to certain constrained scenarios. To this end, we propose a robust general framework for arbitrary biometric matching scenarios without the limitations of alignment as well as the size of inputs. We introduce a feature post-processing step to handle the feature maps from FCN and a dictionary learning based Spatial Feature Reconstruction (SFR) to match different sized feature maps in this work. Moreover, the batch hard triplet loss function is applied to optimize the model. The applicability and effectiveness of the proposed method are demonstrated by the results from experiments on three person re-identification datasets (Market1501, CUHK03, DukeMTMC-reID), two partial person datasets (Partial REID and Partial iLIDS) and two partial face datasets (CASIA-NIR-Distance and Partial LFW), on which state-of-the-art performance is ensured in comparison with several state-of-the-art approaches. The code is released online and can be found on the website: https://github.com/lingxiao-he/Partial-Person-ReID.
... However, the computational cost of each algorithms is extensive and the required alignment step limits its practical applications. Besides, region-based models [5], [12], [25], [28], [29], [33], [34] also offered a solution for partial face recognition. They only required face sub-regions as input, such as eye [33], nose [33], half (left or right portion) of the face [12], or the periocular region [30]. ...
Partial face recognition (PFR) in an unconstrained environment is a very important task, especially in situations where partial face images are likely to be captured due to occlusions, out-of-view, and large viewing angle, e.g., video surveillance and mobile devices. However, little attention has been paid to PFR so far and thus, the problem of recognizing an arbitrary patch of a face image remains largely unsolved. This paper proposes a novel partial face recognition approach, called dynamic feature matching (DFM), which combines fully convolutional networks and sparse representation classification (SRC) to address partial face recognition problem regardless of various face sizes. DFM does not require prior position information of partial faces against a holistic face. By sharing computation, the feature maps are calculated from the entire input image once, which yields a significant speedup. Experimental results demonstrate the effectiveness and advantages of DFM in comparison with state-of-the-art PFR methods on several partial face databases, including CAISA-NIR-Distance, CASIA-NIR-Mobile, and LFW Databases. The performance of DFM is also impressive in partial person re-identification on Partial RE-ID and iLIDS databases. The source code of DFM can be found at
https://github.com/lingxiao-he/dfmnew
.
... Feature set matching [7] has been a hot topic in pattern recognition. [24] was the first work that used graph matching for face recognition. However, their work relies heavily on manual landmarks labeling. ...
... In order to solve partial face recognition, some pioneer works have been proposed. Some approaches [5] [13] [16] use only certain patches of holistic faces, such as the left or right face, periocular, mouth, nose etc. [2] [10] [12] [14] divide the aligned face image into several patches and fuse the matching results by patch-to-patch matching. Nevertheless, it is difficult to detect and align facial portions in practice scenarios when partial face randomly appears. ...
