Citation Context

...during the indoor sessions, and using the frontal camera with 8mm focal length are considered. ROIs have been extracted using the OpenCV implementation of Viola-Jones face and eye detection algorithm =-=[33]-=-. In the same way than with DIEE, faces have been rotated to align the eyes Context-Sensitive Self-Updating for Adaptive Face Recognition 13 Fig. 5 FIA dataset. An example of randomly chosen facial ca...

Citation Context

...ring operations, faces are isolated in the image using the segmentation module, which produces the regions of interest (ROIs). Then, discriminant features are extracted from each ROI (e.g. eigenfaces =-=[32]-=- of local binary patterns [2]) to produce the corresponding pattern d= (d[1], ...,d[F ]) (with F the dimensionality of the feature space). This pattern is then compared to the facial model of each enr...

Citation Context

...e been investigated in the literature over the past decade. They can be organized into the following two categories: 1. Development of discriminative features that are robust to environmental changes =-=[4, 36]-=-. These techniques usually aim to develop facial descriptors insensitive to changes in capture conditions, to mitigate their effects on the recognition process. 2. Storage (or synthetic generation) of...

Citation Context

... of the facial models using the operational data [22, 23]. Adaptation can be either supervised or unsupervised, depending on the labelling process of the operational data. In semi-supervised learning =-=[37]-=-, the facial model of each individual enrolled to the system is updated using operational data labelled as the same individual by the classification system. For example, a gallery Gi of reference patt...

Citation Context

...a product of the three factors – loss of correlation, luminance distortion and contrast distortion: Q(Ri, j,Dl) = σRi, j ,Dl σRi, j ·σDl · 2R̄i, j · D̄l R̄2i, j+ D̄ 2 l · 2σRi, j ·σDl σ2Ri, j +σ 2 Dl =-=(2)-=- where R̄i, j and D̄l are the average images, σRi, j and σDl their variances, and σRi, j ,Dl the covariance. To accommodate spatial variations in image distortion, statistical features for Eq. 2 may b...

Citation Context

...es for adaptive FR using selfupdating is presented, along with the key challenges facing these systems. An experimental protocol involving three real-life facial datasets (DIEE [26], FIA [9] and FRGC =-=[21]-=-) is proposed to evaluate the benefits and drawbacks of a self-updating methodology applied to a template matching system, with a particular focus on the management of system complexity. To address th...

Citation Context

...perational pattern dwould be computed from distance measures to these patterns. Classification may also be performed using neural networks (e.g multi-layer perceptrons [27] and ARTMAP neural networks =-=[3]-=-) or statistical classifiers (e.g. nave Bayes classification [7]), in which case the facial models would consist of parameters estimated during their training using the reference patterns ri, j (e.g. ...

Citation Context

...these patterns. Classification may also be performed using neural networks (e.g multi-layer perceptrons [27] and ARTMAP neural networks [3]) or statistical classifiers (e.g. nave Bayes classification =-=[7]-=-), in which case the facial models would consist of parameters estimated during their training using the reference patterns ri, j (e.g. neural networks weights, statistical distribution parameters, et...

Citation Context

... case matching scores for each operational pattern dwould be computed from distance measures to these patterns. Classification may also be performed using neural networks (e.g multi-layer perceptrons =-=[27]-=- and ARTMAP neural networks [3]) or statistical classifiers (e.g. nave Bayes classification [7]), in which case the facial models would consist of parameters estimated during their training using the ...

Citation Context

...In addition, facial model corruption due to the addition of misclassified templates in the galleries is presented as the ratio of impostor over genuine templates. Following Doddingtons classification =-=[6]-=-, only the 10 galleries with the highest ratio are presented, to focus on lamb-type individuals which are easy to imitate. Finally, a constraint of f pr= 5% has been chosen to compute the decision thr...

Citation Context

... quality index Q [34] is considered to measure the distortion between Dl and each reference ROI Ri, j. This measure is a particular case of the Structural Similarity Index Measure (SSIM) presented in =-=[35]-=-. It can be written as a product of the three factors – loss of correlation, luminance distortion and contrast distortion: Q(Ri, j,Dl) = σRi, j ,Dl σRi, j ·σDl · 2R̄i, j · D̄l R̄2i, j+ D̄ 2 l · 2σRi, ...

Citation Context

...igate their effects on the recognition process. 2. Storage (or synthetic generation) of multiple reference images to cover the different capture conditions that could be encountered during operations =-=[11, 13]-=-. However, these approaches assume that FR is a stationary process, as they only rely on information available during enrolment sessions. In addition, depending on the application, a single enrolment ...

Citation Context

...d being the distance between the eyes), with the eyes located at the position (d/2,d/4) and (3 ·d/2,d/4). In this process, faces have been rotated to align the eyes to minimize intra-class variations =-=[10]-=-, and then normalized to a size of 70x70 pixels. Fig. 4 DIEE dataset. An example of randomly chosen facial captures for two individuals. This dataset was explicitly collected to evaluate the performan...

Citation Context

...dalities. They usually either: 1) add novel captures to individual specific galleries [29], or 2), fuse new input data into common templates referred to as super-templates, containing all information =-=[12, 30]-=- for each modality (for example, virtual facial captures constructed with patches from operational data). This chapter focuses on self-updating techniques with template matching systems for FR. These ...

Citation Context

...ritical [23]. To prevent a decline in classification performance, the use of a strict updating threshold may enable to reduce the probability of updating the facial models with misclassified patterns =-=[14, 29, 30]-=-. However, it has been argued that updating with only highly confident patterns may result in the addition of redundant information in the galleries, and thus a marginal gain in performance at the exp...

Citation Context

...nuous authentication). For example, in the context of controlled access to critical information on computer network systems, the face modality may allow for a continuous, non-intrusive authentication =-=[18]-=-. After initial log-in, a FR system may enroll the authenticated user using facial images captured from the computer’s built-in camera, and design a facial model1. The user’s identity may then be peri...

Citation Context

...r a statistical model estimated from reference captures (statistical classification). 2 A concept can be defined as the underlying data distribution of the problem under specific operating conditions =-=[1]-=-. Context-Sensitive Self-Updating for Adaptive Face Recognition 3 tem operations. Common approaches in adaptive biometrics fall under self-updating and co-updating, depending on whether they rely on a...

Citation Context

...o the reference ROIs Ri, j ( j= 1, ...,Ji) stored in the galleries, and asses whether the capture conditions are novel enough to justify an increase in complexity. The universal image quality index Q =-=[34]-=- is considered to measure the distortion between Dl and each reference ROI Ri, j. This measure is a particular case of the Structural Similarity Index Measure (SSIM) presented in [35]. It can be writt...

Citation Context

...ive biometrics fall under self-updating and co-updating, depending on whether they rely on a single or multiple modalities. They usually either: 1) add novel captures to individual specific galleries =-=[29]-=-, or 2), fuse new input data into common templates referred to as super-templates, containing all information [12, 30] for each modality (for example, virtual facial captures constructed with patches ...

Citation Context

...hting conditions and ageing). To address this limitation, adaptive biometric systems have been proposed in the literature [28], inspired by semi-supervised learning techniques for pattern recognition =-=[17]-=-. These systems are able to adapt facial models (sets of templates or classifier parameters) by exploiting (either on-line or off-line) faces captured during sys1 Depending on the classification syste...

Citation Context

...These methods update template galleries using faces captured during operations that are considered highly-confident, i.e. that produce very high matching scores (surpassing a self-updating threshold) =-=[23]-=-. Advantages and drawbacks of self-updating have been widely investigated [15, 23]. While these methods have been show to significantly improve the performance of biometric systems over time, an updat...

Citation Context

...s they only rely on information available during enrolment sessions. In addition, depending on the application, a single enrolment session is often considered as multiple ones are not always possible =-=[22]-=-. This prevents to integrate new concepts2 that may emerge during operations as capture conditions and individuals physiology evolve over time (for example due to natural lighting conditions and agein...

Citation Context

...ns and individuals physiology evolve over time (for example due to natural lighting conditions and ageing). To address this limitation, adaptive biometric systems have been proposed in the literature =-=[28]-=-, inspired by semi-supervised learning techniques for pattern recognition [17]. These systems are able to adapt facial models (sets of templates or classifier parameters) by exploiting (either on-line...

Citation Context

...dalities. They usually either: 1) add novel captures to individual specific galleries [29], or 2), fuse new input data into common templates referred to as super-templates, containing all information =-=[12, 30]-=- for each modality (for example, virtual facial captures constructed with patches from operational data). This chapter focuses on self-updating techniques with template matching systems for FR. These ...

Citation Context

... necessarily improve performance, and also eventually reduce its response time during operations. To bound this complexity, template management methods (e.g. pruning) have been proposed in literature =-=[8, 15, 23]-=-. While clustering-based methods showed the most promising results, they remain computationally complex and thus not suited for seamless operations, if self-updating is performed frequently. In this c...

Citation Context

...template matching that allows for self-update. 2.3.1 General Presentation To illustrate this principle, it is applied to a template matching system, presented in Figure 2. In this system, inspired by =-=[25]-=-, the facial model of each individual i is designed by storing initial reference patterns from a labelled dataset into a gallery Gi = {ri,1,ri,2, ...} (in this case, the terms pattern and template are...

Citation Context

...reover, such systems are usually initialized with a limited number of high-quality reference face captures, which may generate non-representative facial models (not modelling all possible variations) =-=[19]-=-. To account for such intra-class variations, several solutions have been investigated in the literature over the past decade. They can be organized into the following two categories: 1. Development o...

Citation Context

...g threshold, preventing the classification system to assimilate this new information. To address this limitation, coupdating methods have been proposed to benefit from complementary biometric systems =-=[23, 24]-=-. Each system is initialized with reference templates from a different source (or different features extracted from the same source), and performs classification of operational input data. In the same...

Citation Context

...known that the performance of state-of-the-art FR systems may be severely affected by changes in capture conditions (e.g., variation in illumination, pose and scale), as well as individual physiology =-=[20, 31]-=-. Moreover, such systems are usually initialized with a limited number of high-quality reference face captures, which may generate non-representative facial models (not modelling all possible variatio...

Citation Context

...e been investigated in the literature over the past decade. They can be organized into the following two categories: 1. Development of discriminative features that are robust to environmental changes =-=[4, 36]-=-. These techniques usually aim to develop facial descriptors insensitive to changes in capture conditions, to mitigate their effects on the recognition process. 2. Storage (or synthetic generation) of...

Citation Context

... that are considered highly-confident, i.e. that produce very high matching scores (surpassing a self-updating threshold) [23]. Advantages and drawbacks of self-updating have been widely investigated =-=[15, 23]-=-. While these methods have been show to significantly improve the performance of biometric systems over time, an updating strategy only relying on matching score values may add redundant template to t...

Citation Context

...ource is not necessarily observed on others. A recent model has been proposed to estimate optimal amounts of samples and iterations to improve system’s performance under specific updating constraints =-=[5]-=-. This model has shown to be effective under the stringent hypothesis of 0% false alarm rate for the updating threshold of both systems. While co-updating is usually applied with multiple biometric tr...

Citation Context

...a period of 1.5 years. For simulations, the facial captures or each individuals are separated into 6 batches corresponding to the capture sessions. ROIs have been extracted with a semi-manual process =-=[16]-=-: an operator first selected the eyes in each frame, and the cropped region was then determined as the square of size 2d ∗ 2d (d being the distance between the eyes), with the eyes located at the posi...

Citation Context

...known that the performance of state-of-the-art FR systems may be severely affected by changes in capture conditions (e.g., variation in illumination, pose and scale), as well as individual physiology =-=[20, 31]-=-. Moreover, such systems are usually initialized with a limited number of high-quality reference face captures, which may generate non-representative facial models (not modelling all possible variatio...

Citation Context

...ate-of-the-art techniques for adaptive FR using selfupdating is presented, along with the key challenges facing these systems. An experimental protocol involving three real-life facial datasets (DIEE =-=[26]-=-, FIA [9] and FRGC [21]) is proposed to evaluate the benefits and drawbacks of a self-updating methodology applied to a template matching system, with a particular focus on the management of system co...

Citation Context

...-art techniques for adaptive FR using selfupdating is presented, along with the key challenges facing these systems. An experimental protocol involving three real-life facial datasets (DIEE [26], FIA =-=[9]-=- and FRGC [21]) is proposed to evaluate the benefits and drawbacks of a self-updating methodology applied to a template matching system, with a particular focus on the management of system complexity....

Citation Context

...igate their effects on the recognition process. 2. Storage (or synthetic generation) of multiple reference images to cover the different capture conditions that could be encountered during operations =-=[11, 13]-=-. However, these approaches assume that FR is a stationary process, as they only rely on information available during enrolment sessions. In addition, depending on the application, a single enrolment ...