A Novel Image Fusion Scheme For Robust Multiple Face Recognition With Light-field Camera

Transcription

1 A Novel Image Fusion Scheme For Robust Multiple Face Recognition With Light-field Camera R. Raghavendra Kiran B Raja Bian Yang Christoph Busch Norwegian Biometric Laboratory, Gjøvik University College, Norway. The Norwegian Colour and Visual Computing Laboratory, Gjøvik University College, Norway. raghavendra.ramachandra@hig.no Abstract Accurate face recognition in wide area surveillance application is a challenging problem because of multiple pose, non-uniform illumination, low resolution and out-of-focus face images that are recorded with conventional surveillance cameras (Closed-Circuit TeleVision or Pan-Tilt-Zoom). In this paper, we address the problem of face recognition in wide area surveillance with a light-field camera. The main advantage of a light-field camera is that, it can provide different focus (or depth) images in a single exposure (capture) which is not possible with a conventional 2D camera. In this work, we propose a novel weighted image fusion scheme to combine different depth (or focus) images rendered by a light-field camera. The proposed image fusion scheme is not only dynamic in handling number of depth (or focus) images but also adaptive in assigning higher weights to the best focus image as compared to the out-offocus image. Extensive experiments are carried out on our newly acquired face dataset captured using Lytro light-field camera to bring out the merits and demerits of the proposed weighted image fusion scheme for face recognition in wide area surveillance applications. I. INTRODUCTION The face detection and recognition has witnessed a rapid growth from past two decades and is also one of the well addressed research areas in the discipline of computer vision. Even after considerable amount of work is carried out to achieve the acceptable accuracy, face recognition still remains as a challenging problem, especially in the presence of varying illumination, multiple poses, at a distance, etc. Since the performance of the face recognition system strongly depends on the acquisition device (or camera) used to capture the face data, the use of particular cameras (like CCTV, PTZ) will also impact the overall performance of the face based biometric system. In general, cameras employed for real-life wide area surveillance exhibit a distinct perspective and have fixed focus that frequently result in out-of-focus face samples, which causes degraded performance of a face recognition system. One possible way to address this problem is by employing a Light-Field Camera (LFC). The light field camera allows to capture the ray-space which holds the rich information such as color, intensity, position and direction of incoming rays of light. The light field camera captures the image by sampling the4d light-field on its sensor in a single photographic exposure by inserting a microlens array [5] or a pin-hole array [4] or masks [18] between the sensor and the main lens. Thus, the presence of the micro lens (or array of pin-holes or masks) allows one to measure not just the total amount of light intensity deposited on the sensor, but also the direction of each ray of incoming light. Finally, by re-sorting the measured rays of light to where they would have terminated one can obtain a number of sharp images focused at different depth [5]. Thus, the light-field camera exhibits interesting features as compared to a conventional camera, such as (1) generating images at different focus (or depth) in one shot. (2) it is low cost; it does not require to move the lens to set the focus on the object in a scene. (3) portable and hand-held (4) real-time exposure. Thus, these features of LFC motivates us to investigate its applicability for the wide area surveillance that involves identifying multiple faces with different pose that are present at various distance (between 0.5m 20m). We study this by simulating real-life surveillance scenarios. There exists a number of weighted fusion schemes especially for multimodal biometric system in which information from different modalities are combined at four different levels, namely: [14]: sensor [17] [11], feature [10] [9], comparison score [12] [20] [15] and decision level [16]. The available work on weighted fusion schemes can be broadly classified into two types: (1) Multimodal weighted fusion: Here, the weight assignment is carried out based on the performance (measured in terms of Equal Error Rate (EER) or/and False Acceptance Rate (FAR) and/or False Reject Rate (FRR)) of the individual modalities to be combined. Thus, higher weight is assigned to the modality that exhibits the best performance over other modalities. (2) Unimodal multisensorial weighted fusion: Here, the unimodal biometric modality acquired with different sensors (for example, face acquired using visible and thermal sensors) [11] or with a same sensor under different acquisition conditions, for example, multispectral acquisition of palmprints [22] are fused. In most of the cases, the weighted fusion is normally carried out at sensor/image level. Here, the weight assessment is carried out either at pixel level or at image level [11] [22]. In case of pixel level weight assessment, each pixel is weighted according to its contribution to the overall biometric system performance. Here, main idea is to use the optimization schemes like Genetic algorithms [3], or particle swarm optimization (PSO) [11] to accurately compute the weight for each pixel in an iterative manner till the performance gain is achieved. In case of image level weighting scheme, the wavelet transform can be used to decompose the images to be fused, then, these decomposed wavelet coefficients are analyzed to perform either selection or fusion (usually equal weighted). Finally, based on these selected or fused wavelet coefficients are used to obtain a fused image [22] [13] using Inverse wavelet transform. In this work, we propose a novel weight assessment

2 Light-field data acquisition using Lytro camera Depth Image 1 Depth Image N Face detector 1 Face detector N Proposed selection and Weighted image fusion Feature Extraction and Identification Fig. 1. Block diagram of the proposed scheme (a) Fig. 2. Lytro light-field Depth images (a) Different depth images (scaled to fit the page) Segmented faces using face detector scheme to accurately combine different focus (or depth) images rendered by the Lytro light field camera. Even though the proposed method belongs to the class of unimodal image fusion (face modality in our case), it still stands out in two different aspects: (1) Since, the number of depth (or focus) images rendered by the Lytro light-field camera is not constant, thus, the proposed scheme must be dynamic to incorporate this property of the Lytro light field camera, (2) Since, we are addressing the multiple face recognition in a given scene, each subject will have their best focus images in any one of the focus (or depth) images and hence, the proposed methods should be adaptive i.e. it should assign larger value of weight for good focus images as compared to that of the out-offocus image. To address this, we propose a novel weighted image fusion scheme that can be structured in two steps: (1) Image selection based on wavelet entropy (2) Novel weight assessment on each of these selected images before performing the fusion using weighted sum rule. In order to effectively evaluate the proposed schemes, we constructed a new multiple face dataset by simulating real life surveillance scenario. To this extent, we employed the first available consumer LFC developed and marketed by Lytro Inc. [1]. We then designed three different (indoor, corridor and outdoor) image acquisition protocols reflecting real life surveillance scenarios. We acquired a new face dataset comprising of 25 subjects such that, each is having 8 enrollment samples acquired in the studio setting using Canon EOS 550D and there are 303 probe samples acquired using Lytro LFC [1]. The rest of the paper is organized as follows: Section II will discuss the proposed method, Section III presents the details on dataset collection and results. Section IV draws the conclusion. II. THE PROPOSED SCHEME Figure 1 shows the block diagram of the proposed method that can be structured in three steps: (1) Face detection and extraction (2) Proposed image selection and weighted fusion scheme (3) Feature extraction and identification. The following section describes these steps in detail. A. Face detection and extraction The first step involves extracting the face region. In this work, we employed the well known Viola-Jones face detector [19] by considering its robustness and performance in a reallife scenario. We trained the face detector using 2429 face samples and 3000 non-face samples acquired using a conventional camera. Figure 2 illustrates the qualitative results of the face detector (Figure 2 ) on different depth images rendered by Lytro LFC. Each image acquired using Lytro LFC will result in multiple depth images. Since, each of these depth images will have at-least one region in focus, applying face detector on any one of these depth images may not result in

3 accurate face region detection (because of out-of-focus face regions). Hence, in this work, we carry out face detection on each of these depth images and then select the result of face detection that corresponds to maximum number of detected face region. We then use this as a reference image and extract the corresponding face region from all depth images. B. Proposed image selection and weighted fusion scheme Depth Image 1 Depth Image 2 Depth Image N DWT DWT DWT Calculate Entropy (E 1 ) Calculate Entropy (E 2 ) Image selection based on positive entropy (I s1, I s2,..i sn ) Calculate Entropy (E N ) In the next step, we propose to perform the image selection by measuring the quality (or focus) by computing the entropy of the wavelet coefficients (W 1 ). In this work, we are motivated to use the entropy as a measure of focus (or quality) of the image (1) as it can provide monotonic quantitative values i.e., the higher values are noted for best focused images while low values are noted for out-of-focus images. Thus, the higher values always correspond to good quality images while lower values represent the low quality images. (2) because of its robustness to noise. These features make the wavelet entropy measure more suitable to evaluate the focus (or quality) of the image. Given the wavelet coefficients W 1, the log-energy entropy can be obtained as follows [8]: E 1 = K (log 2 (W 1i ) 2 ) (1) i=1 Where, K denotes the number of wavelet coefficients obtained on an image (say I 1 (x,y)) and E 1 denotes the log-energy entropy of the corresponding image I 1 (x,y). Normalize and sorting (S 1, S 2,..,S j ) Sliding window based differencing (D 1, D 2,..,D j-1 ) E 1 = 4535 E 2 = 7574 E 3 = 9243 E 4 = 6393 E 5 = -556 E 6 = E 7 = (a) Weights computation (w 1, w 2,..,w j ) IDWT Weighted fusion Fused Image Nor 1 = Nor 2 = Nor 3 = Nor 4 = Sor 1 = Sor 2 = Sor 3 = Sor 4 = (c) D 1 = D 2 = D 3 = w 1 = w 2 = w 3 = w 4 = Fig. 3. Block diagram of the proposed image selection and weighted fusion scheme Figure 3 shows the block diagram of the proposed image selection and weighted fusion scheme to combine information from different focus (or depth) images rendered by the Lytro light-field camera. The proposed fusion scheme can be broadly viewed in two levels: (1) Image selection based on the wavelet log-energy entropy measure, (2) Weighted fusion of the selected images such that, weights are assigned in an adaptive and dynamic manner. 1) Image selection: The proposed image selection is carried out on focus (or depth) images (face images obtained after face detector) corresponding to one subject at a time. Let N be the number of face images corresponding to different focus (or depth) from the first subject such that: Sub 1 = {I 1 (x,y),i 2 (x,y),...,i N (x,y)}. The first step of the proposed selection method involves in decomposing a face image corresponding to different focus (or depth) using 2D- Discrete Wavelet Transformation (DWT) [7]. Given a focus image I 1 (x,y), the 2D-DWT operation is performed by carrying out two operations namely [8]: filtering and downsampling using low-pass (L) and high-pass (H) filter along both row (x) and column (y) of I 1 (x,y). Let W 1 represent the decomposed wavelet coefficients corresponding to the focus image I 1 (x,y). Fig. 4. scheme (d) Illustration of the proposed image selection and weighted fusion The above procedure is repeated for all N images corresponding to the subject Sub 1 to compute the corresponding entropy E = {E 1,E 2,...,E N }. Figure 4 (a) illustrates this step by showing the different focus images corresponding to the subject (Sub 1 ) and their corresponding entropy values. In the next step, we perform the image selection by choosing all images whose entropy value is greater than 0 (i.e. positive entropy value). Figure 4 illustrates the selected images and let these selected images be denoted as I s = {I s1,i s2,...,i sn }. Thus, it can be observed from Figure 4(a) and that, adoption of the Log-energy based entropy measure [8] can successfully provide the quantitative measure of the focus. 2) Weighted image fusion: After selecting the images based on the wavelet entropy measure, we propose a new method to compute the weight for each of these selected images. The core idea of the proposed weight assessment scheme is to assign high weight to the image with large wavelet entropy value and low weight to the image with lower wavelet entropy

4 value in a dynamic and adaptive fashion. The proposed weight assessment scheme is detailed in the following steps: a) Normalizing and sorting: Given the selected images and their corresponding wavelet entropy E sel = {E s1,e s2,...,e sn }, where, sn represents the number of selected images. We first perform the normalization on the entropy values E sel as follows: Nor j = E j max(e sel ) Where, Nor j j = s1,s2,...,sn represents the normalized entropy values of the selected images, E j represents the wavelet entropy value of the j th image and max(e sel ) indicates the maximum wavelet entropy value of the selected images. In the next step, we sort normalized values in decreasing order, such that, the image with highest wavelet entropy value appears first. (2) Sor = sort{nor 1,Nor 2,...,Nor j } (3) Finally, the normalized and sorted wavelet entropy values can be represented as: Sor = {Sor 1,Sor 2,...,Sor j } and { the corresponding sorted images can be written as I s = I 1,I 2,...,I j }. b) Sliding window difference: Here, we perform the difference between consecutive normalized and sorted entropy values (Sor). The sliding window employed in this work is of dimension 2 with an overlap of 1. Here, the main idea of performing the sliding window difference is to analyze the change in the normalized and sorted entropy values (Sor) which in turn will be used to assign the corresponding weights. Let D be the difference obtained using the sliding window as follows: D j = Sor j+1 Sor j (4) c) Weight Assessment: The proposed weight assessment scheme will dynamically assign the weight on each of the selected focus images. The weight assessment is carried out based on the difference value D j as it can provide the clue on information between the images. Thus, the low value of D j implies that corresponding images (say I 1 and I 2) are of equal importance and hence an equal weight can be assigned. While, higher values of D j implies that both images (say I 1 and I 2) are different and hence require different weights. In this work, we propose a novel function that can dynamically compute the weight by satisfying the above mentioned criteria. It can be formulated as follows: F = 0.5+(0.5 D j ) (5) { F Max Weight, if Th 0.2 w j = M ax W eight/2, otherwise Figure 5 shows the response of the proposed function (Equation 5) for different values of D j. Here, it can be observed that, if the value of D j is less than the threshold (T h), then equal weights are assigned, else, different weights are assigned for the image. Since, images are sorted according (6) Fig. 5. Response of the proposed weight mapping function to the largest entropy value, the proposed method will always assign highest weight to the image that has a large entropy value. The complete weight assessment scheme is illustrated in the Figure 4 (c) and the whole procedure is summarized in the Algorithm 1. The value of Th = 0.2 is chosen based on different trials by considering the overall performance of the proposed scheme. Algorithm 1 Proposed algorithm for weight assessment Max Weight 1; Th 0.2; if d j < Th then % Assign equal weights w j Max Weight/2; w j+1 Max Weight w j ; Max Weight w j+1 ; % Update Max Weight else % Assign different weights F 0.5+(0.5 D j ); w j Max Weight F; w j+1 Max Weight w j ; Max Weight w j+1 ; % Update Max Weight end if d) Weighted image fusion: After accurately computing the weight for each image, we carry out the image fusion using weighted SUM rule as follows: Im Fu = j I j w j (7) where, Im Fu represents the fused image, I j (prime) represents the j th image and w j represents the computed weight for the j th image such that, w 1 +w 2 +w w j = 1. Finally the Inverse Discrete Wavelet Transform (IDWT) is employed to reconstruct the fused image on which the feature extraction and identification is carried out. Figure 4 (d) shows the fused image obtained using proposed image selection and fusion scheme. Figure 9 shows the results of the proposed weighted image fusion method on different subjects. The Figure 9 (a) shows images obtained after performing the selection based on positive entropy values and Figure 9 shows the results of the proposed weighted image fusion scheme. Thus, it can be observed that, the fused images obtained using the proposed scheme appears to have superior visual details as compared to the selected images of Figure 9 (a).

5 (a) Fig. 6. Example of the enrolled sample corresponding to one subject: (a) Enrolled samples Corresponding face images (a) (c) Fig. 7. Examples of probe samples: (a) Corridor Indoor (c) Outdoor (for simplicity, results are shown on one of the possible depth images) (a) (c) Fig. 8. images) Qualitative results of face detector on (a) Indoor Corridor (c) Outdoor scenario (for simplicity, results are shown on one of the possible depth

6 (a) Fig. 9. Results of the proposed scheme (a) After image selection After weighted image fusion normal lighting conditions (because of fluorescent lamps) that are already present in the corridor. (3) Outdoor: Here, probe samples are acquired in outdoor conditions in which we have naturally varying sunlight. Examples of the probe samples are shown in Figure 7. In all three scenarios, the subjects are allowed to stand between the distance of 0.5m 20m and every probe sample has 2 to 4 subjects standing at various distance with varying poses. With these settings, we captured 303 probe samples that resulted in 986 face samples from 25 subjects that are distributed in three different scenarios like indoor (433), outdoor (340) and corridor (213). A. Results and discussion C. Feature extraction and identification After obtaining the fused image, we perform the feature extraction using two well established algorithms in the field of face biometrics, namely: Local Binary Patterns (LBP) [6] and Log-Gabor filters [2]. The LBP employed in this work has a radius of 2 while Log-Gabor filter has 4 scales and 8 orientations and we fix these values as the result of experimental trials and also in conformity with literature [10]. We then use these extracted features to accurately identify the subject by employing the Sparse Reconstruction Classifier (SRC) [21]. Here, we carry out l1 - minimization via SP GL1 solver based on spectral gradient projection [21]. In this work, we obtain the comparison score that directly corresponds to the residual errors obtained using SRC. III. E XPERIMENTS AND DISCUSSION This section presents the experimental protocols and results of the proposed weighted image fusion scheme on our newly collected multiple face light-field face dataset using Lytro light field camera [1]. The whole dataset was acquired in our laboratory over a period of 20 days and comprises of 25 subjects. The data acquisition process is divided into two parts, namely: Enrollment data samples and probe data samples. The enrollment samples are collected in a controlled illumination (studio) environment using a Canon EOS 550D DSLR camera. For each subject, we captured 8 samples that correspond to neutral, smile and 6 different poses as shown in the figure 6 (a). The acquired enrollment samples are in lossless 24 bit color JPEG format with the original size of pixels. In order to carry out the experiments, we perform the pre-processing steps to accurately extract the face region. Our pre-processing steps include a face detector to detect and crop the face region which is then re-sized to have the size of pixel. Then we finally perform the Gaussian filtering (σ = 2) to remove the noise contained in high spatial frequency. Figure 6 shows the cropped face images that are used for experiments. Thus the enrollment dataset consists of 8 samples per subject that resulted in a total of 200 samples. The probe samples are collected using the Lytro light-field camera in three different scenarios, namely: (1) Indoor: Here, all probe samples are collected in the room with controlled lighting without any sunlight. (2) Corridor: Here, subjects are allowed to stand close to the windows through which sunlight is allowed to fall on the subject s face in addition to the This section discusses the results of the proposed image fusion algorithm on our database collected using Lytro lightfield camera. The experiments are carried out on 303 probe samples that are collected by simulating real-life surveillance scenarios. In this work, all the results are presented in terms of identification rate (rank 1) that is obtained by comparing 1 : N subjects in the dataset, therefore, a higher value of identification rate corresponds to better accuracy of the algorithm. In order to effectively evaluate the face recognition algorithms on our new dataset, we employ the gallery samples from our enrollment dataset and probe samples corresponding to the images acquired using Lytro light-field camera. Figure 8 shows the qualitative results of the face detection algorithm on the probe samples acquired using Lytro lightfield camera. For simplicity, the results on light-field camera are illustrated on one of the depth images. In order to obtain the quantitative results, we evaluate the face detector on every probe sample that shows the detection rate of 94.8%. In this work, we present the results of both image selection and weighted image fusion method discussed in section II. The first proposed scheme is based on the image selection using the entropy based measure and we denote this scheme as the Proposed scheme 1. The second scheme is based on the proposed weighted image fusion and we denote this as the Proposed scheme 2. (a) Fig. 10. Qualitative illustration (a) Best focus image Proposed weighted image fusion

7 Figure 10 shows the qualitative results of the proposed weighted image fusion scheme and the best focus image selected based on the wavelet entropy. It can be observed that the proposed weighted fusion scheme has shown greater visual quality as compared with the best focused image. This fact illustrates the efficacy of the proposed weight assessment scheme. Table I shows the quantitative performance of the proposed scheme 1 and proposed scheme 2 when LBP and SRC is employed as the feature extraction and classification scheme. From the Table I, it can be observed that, the proposed scheme 2 has outperformed the proposed scheme 1 in all three scenarios. The best result is noted for indoor scenario with an identification rate of 58.66%. TABLE I. QUANTITATIVE PERFORMANCE OF THE PROPOSED IMAGE FUSION SCHEMES: LBP-SRC Method Scenario Identification Rate (%) Proposed Scheme 1 Proposed scheme 2 Indoor Corridor Outdoor Indoor Corridor Outdoor Table III shows the quantitative performance of the decision level fusion of LBP-SRC and LG-SRC using OR rule. It can be noted that the best performance is obtained with the proposed scheme 2 for corridor scenario with an identification rate of 75.12%. Thus, from the above experiments, it can be observed that the proposed scheme 2 has shown the best performance in all three different scenarios and thereby justifying the efficacy. IV. CONCLUSION In this paper, we proposed a novel weight assessment scheme to carry out the weighted image fusion scheme to combine different depth (or focus) images for accurate face recognition. The proposed weight assessment scheme is not only dynamic but also adaptive in assigning the weights for different depth (or focus) image rendered by Lytro lightfield camera. The proposed scheme is evaluated on the newly collected multiple face dataset consisting of 25 subjects in three different scenarios resulting in 303 probe samples with 986 face samples scattered between the distance of 0.5m - 20m. Experimental results have revealed that, the proposed scheme 2 has shown the performance with an identification rate of 75.12%. Table II shows the qualitative performance of the proposed scheme I and II when LG-SRC combination is used as the feature extraction and classification method. Here, it can also be observed that, the proposed scheme 2 outperforms the proposed scheme 1 in all three different scenarios. Here, the best performance is noted for the proposed scheme 2 with an identification rate of 54.93% In order to further TABLE II. QUANTITATIVE PERFORMANCE OF THE PROPOSED IMAGE FUSION SCHEMES: LG-SRC Method Scenario Identification Rate (%) Proposed Scheme 1 Proposed Scheme 2 Indoor Corridor Outdoor Indoor Corridor Outdoor improve the overall performance of the system by exploiting the complementary information on the decision provided by each of these algorithms (LBP-SRC and LG-SRC), we propose to carry out the multi-algorithm fusion at the decision level. In this work, we employed the OR rule to combine the decisions from multiple algorithms. TABLE III. QUANTITATIVE PERFORMANCE OF THE PROPOSED IMAGE FUSION SCHEMES: DECISION LEVEL FUSION Method Scenario Identification Rate (%) Proposed scheme 1 Proposed scheme 2 Indoor Corridor Outdoor Indoor Corridor Outdoor V. ACKNOWLEDGMENT This work is funded by the EU 7th Framework Program (FP7/ ) under grant agreement n o for the large-scale integrated project FIDELITY. REFERENCES [1] Homepage of lytro:-. [2] F. J. David. Relations between the statistics of natural images and the response properties of cortical cells. Journal of the Optical Society of America, 4(12): , [3] G.Bebis, A.Gyaourova, S.Singh, and I.Pavlides. Face recognition by fusing thermal and visible imagery. Image and Vision Computing, 24: , [4] R. Horstmeyer, G. Euliss, R. Athale, and M. Levoy. Flexible multimodal camera using a light field architecture. In International Conference on Computational Photography (ICCP), pages 1 8, [5] R. Ng, M. Levoy, M. Bredif, G. Duval, M. Horowitz, and P. Hanrahan. Light field photography with a hand- held plenoptic camera. In Stanford University Computer Science Technical Report CSTR , pages 1 11, [6] T. Ojala, M. Pietikainen, and T. Maenpaa. Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 24(7): , [7] G. Pajares and J. M. de la Cruz. A wavelet-based image fusion tutorial. Pattern Recognition, 37(9): , [8] G. Pajares and J. M. de la Cruz. A wavelet-based image fusion tutorial. Pattern Recognition, 37(9): , [9] R. Raghavendra. PSO based framework for weighted feature level fusion of face and palmprint. In Eighth International Conference on Intelligent Information Hiding and Multimedia Signal Processing (IIH- MSP), pages , [10] R. Raghavendra, B. Dorizzi, A. Rao, and G. H. Kumar. Designing efficient fusion schemes for multimodal biometric system using face and palmprint. Pattern Recognition, 44(5): , [11] R. Raghavendra, B. Dorizzi, A. Rao, and G. H. Kumar. Particle swarm optimization based fusion of near infrared and visible images for improved face verification. Pattern Recognition, 44(2): , [12] R. Raghavendra, A. Rao, and G. H. Kumar. Qualitative weight assignment for multimodal biometric fusion. In 7th International Conference on Advances in Pattern Recognition (ICAPR 2009). Kolkata, India, pages , 2009.

8 [13] R. Raghavendra, A. Rao, and G. H. Kumar. Multisensor biometric evidence fusion of face and palmprint for person authentication using particle swarm optimisation (pso). International Journal of Biometrics, 2(1):19 33, [14] Ross.A, Nandakumar.K, and J. A.K. Handbook of Multibiometrics. Springer-Verlag edition, [15] Y. N. Singh and P. Gupta. Qualitative Evaluation of Normalization Techniques of Matching Scores in Multimodal Biometrics Systems. In proceedings of International Conference on Biometrics (ICB), pages , [16] S.Prabhakar and A. Jain. Decision level fusion in fingerprint verification. Pattern Recognition, 35(4): , [17] S.Singh, A.Gyaourova, G.Bebis, and I.Pavlidies. Infrared and visible image fusion for face recognition. In of SPIE Defense and security symposium, pages , [18] A. Veeraraghavan, R. Raskar, A. Agrawal, A. Mohan, and J. Tumblin. Dappled photography: mask enhanced cameras for heterodyned light fields and coded aperture refocusing. ACM Trans. Graph., 26(3):1 12, July [19] P. Viola and M. Jones. Robust real-time face detection. International Journal of Computer Vision, 57: , [20] Y. Wang, T. Tan, and A. K. Jain. Combining Face and Iris Biometrics for Identity Verification. In proceedings of 4th International Conference on Audio and Video based Biometric Person Authentication (AVBPA, Guildford, UK), pages , [21] J. Wright, A. Y. Yang, A. Ganesh, S. S. Sastry, and Y. Ma. Robust face recognition via sparse representation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(2): , [22] Z. S. Y. Hao and T. Tan. Comparative studies on multispectral palm image fusion for biometrics. In International Conference on Asian Conference on Computer Vision, page 1221, 2007.