A Novel Image Deblurring Method to Improve Iris Recognition Accuracy

Transcription

1 A Novel Image Deblurring Method to Improve Iris Recognition Accuracy Jing Liu University of Science and Technology of China National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences Zhenan Sun National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences Tieniu Tan National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences Abstract Iris recognition provides a reliable method for personal identification. However, it is inevitable that a large portion of iris images captured in less-intrusive systems are out-offocus or motion blurred. Blurred iris images without clear iris texture details will lead to high false reject rate of iris recognition. In this paper, a novel image deblurring method is proposed to automatically enhance the quality of both defocused and motion blurred iris images, and it mainly includes three steps. Firstly, each blurred iris image is classified into two categories, i.e. defocused or motion blurred sample. Secondly, the initial point spread function (PSF) is automatically refined based on selected gradient maps after a few iterations. Finally, a step of image deconvolution is performed by adopting a more accurate noise model. Experimental results on various iris image databases illustrate that the proposed method is effective and outperforms stateof-the-art image deblurring methods in the improvement of iris recognition accuracy. 1. Introduction Iris recognition [5] becomes a hot topic in biometrics recently because of its advantages in large-scale and accurate identity authentication. Current iris recognition systems perform well under relatively controlled environments, while the design of methods that are effective in lessintrusive system is still a major area of research. As shown in Figure 1, blur is inevitably present in practical applications and is usually originated from out-of-fucus, motion of imaging objects and sometimes both of them. So the user has to hold still in case of motion blur and to be a specific distance from the camera in case of defocus blur. In order to relax the strict requirements of position and motion for users and improve iris recognition accuracy, it is necessary to develop an image deblurring method to enhance the quality of blurred iris images. For the images captured in nature scenes, state-of-the-art image deblurring methods [3, 7, 14] all estimate PSF and latent unblurred image simultaneously to deal with complicated PSF. These methods predict the latent unblurred image and use a number of iterations to solve the optimization problem. Because of the usage of prior information, those algorithms are complicated and the computational cost is very high. For iris deblurring methods, most of them make use of the characteristics of iris image [10, 11] or extra hardware [9]. To our best knowledge, those methods are based on the assumption that the PSF has a known form and only its parameters are required to be estimated from the iris image, so they are only effective in one special category of blurred iris images, either defocus [9, 10] or motion blur [11]. In additional, the PSF with assumed form is not precise enough to deconvolve the image without degradation. In this paper, an automatic Iris image Deblurring method with PSF Refinement (IDPR) is proposed to improve iris recognition accuracy without any extra information of sensors, depth information or inter-frame information. The proposed method can be applied for both defocused and motion blurred samples, so it is advantageous over pervious methods which only work in a single blurring condition. In addition, the initialized PSF in IDPR get precise after a few iterations of refinements. The contributions of this paper are summarized as follows. 1) A novel iris deblurring model including three modules is proposed to accomplish the restoration of blurred iris images. Unlike conventional methods which just initialize the PSF to restore a blurred iris image, this model consists of PSF initialization, PSF refine /11/$ IEEE

2 Figure 1. Examples of clear, defocused, motion blurred, defocused and motion blurred images (from left to right). ment and final deconvolution. 2) An efficient approach is proposed to refine PSF. Instead of using the original image directly like other conventional methods, the selected gradient maps are applied in conjunction with a noise model describing the randomness precisely to refine the PSF efficiently. 3) Utilizing the characteristics of iris image, we propose a whole set of methods to initialize PSF, three rules to select gradient maps and a new regularization term in PSF refinement, make IDPR robust and efficient. Extensive experiments show that IDPR can improve accuracy and robustness of iris recognition systems and enlarge the effective capture range of iris recognition systems. The remainder of this paper is organized as follows: In Section 2, overview of the proposed method is presented. Section 3 and Section 4 describe the details of two parts of the proposed method respectively. The experimental results are provided in Section 5. Section 6 concludes this paper. 2. Iris image deblurring method with PSF refinement 2.1. Background In general, the shift-invariant blurring is always modeled as a convolution process, I = L f + n, (1) where I, L, and n represent the degraded image, latent unblurred image, and additive noise respectively. denotes the convolution operator, and f is a shift-invariant PSF. The restoration methods of blurred images can be divided into two classes. The first one where PSF is known and only L is required to be estimated, is called non-blind deconvolution. In the other one, PSF and latent unblurred image are estimated simultaneously. It is a blind deconvolution process which is severely ill-posed. Our method is developed from the later class. Solving a blind deconvolution problem is usually formulated in a Bayesian framework [14], p(l, f I) p(i L, f )p(l)p(f ), (2) where p(i L, f ) is the likelihood, and p(l),p(f ) denote the priors on the unblurred image and the PSF respectively. This maximum a posteriori (MAP) problem can be transformed into an energy minimization problem that minimizes the negative logarithm of Equation (2) with a Gaussian assumption. Then, the energy minimization problem can be written as, E(L, f ) = L f L 2 + αψ(f ) + βφ(l), (3) where Φ(L),Ψ(f ) are regularization terms for L and f. To solve the energy minimization problem, Equation (3) is divided into two sub-problems, namely PSF estimation and L estimation, fˆ = arg min L f L 2 + αψ(f ), (4) L = arg min L f L 2 + βφ(l). (5) f L They are optimized alternatively. Equation (5) is also the energy function for non-blind deconvolution Structure of IDPR In the previous iris image deblurring methods, the form of PSF is supposed to be known and only the parameters need to be estimated. After the PSF is obtained by estimating the parameters, the problem becomes a non-blind deconvolution. PSF cannot reflect the actual blurring factor once its assumed form is not fitting for the actual blurring reason. To address it, IDPR adopts a novel algorithm structure which is developed from blind deconvolution, and makes use of the characteristics of iris images. From the flow chart of IDPR in Figure 2, we can see that IDPR can be separated into three parts. Firstly, a whole set of methods to initialize PSF is proposed, and it can obtain the initialized PSF for both motion and defocus blurring. For each category, there is a sub-unit to estimate the parameters of the initial PSF. Secondly, PSF refinement consists of PSF estimation and L estimation which are optimized alternatively. To be efficiently, the selected gradient information is applied instead of the pixel s intensity value which is commonly used in other methods, and an accurate noise model is adopt in PSF and L estimation. Taking into account the characteristic of iris images, we propose three rules to select the effective regions of gradient maps and apply the initial PSF as a regularization term

3 Figure 3. The original image (left) and the result of iris segmentation (right). Figure 2. The flow chart of the proposed method. in PSF estimation for robustness and efficiency. Finally, L estimation is performed using the refined PSF to obtain a deblurred image. Based on the accurate PSF which can reflect the actual blurred reason, the image deconvolution can get a satisfying result. 3. PSF initialization Because high-frequency components both in out-offocus and motion blurred images are lost, it is difficult to determine the category of a blurred iris image. And it is more challenging to obtain the weight of each blurring factor for a mixed defocused and motion blurred image. In IDPR, a set of PSF initialization methods is proposed to address the above problem. The first step of our algorithm is to analyze the main factor for a blurred image. With the assistant of two-field C- CD [2], a very different scene will be present by the two fields in a motion blurred image. As Wei et al. proposed in [17], the difference between two adjacent rows can be adopted as a measurement to determine which following sub-unit is used. Afterwards, one field is extracted for subsequent processing Defocus blurred iris image In a general way, the defocus blurring PSF is modeled as a Gaussian kernel, f(x, y) = 1 2πσh 2 e x 2 +y 2 2σ h 2. (6) where σh 2 represents the variance and is difficult to be estimated precisely [16, 10]. Thanks to the PSF refinement, only an approximate PSF is required in this step. First the circular boundaries of an iris are localized [4], and then the non-iris region is excluded [8] as shown in Figure 3. The focus value (FV) proposed by Daugman [6] is measured on the detected iris region without the influence of eyelashes.to obtain the relationship between σ h and FV, 10 clear iris images are blurred with different σ h values artificially. From the calculated FVs of these images and their σ h value, a relation curve with least fitted error can be obtained. According to this curve, σ h can be calculated using the focus value Motion blurred iris image Usually, the PSF of linear motion blur is modeled as, 1 d if 0 x d cos( θ + π 2 ) f(x, y) = and 0 y d sin( θ + π 2 ), (7) 0 otherwise where θ and d represent the direction and length of motion respectively. As shown in Figure 1, the iris images always have specular reflections on cornea surface. Moreover, the length of the specular reflection is proportional to the motion distance. Thus, the specular reflection is applied to estimate the parameters of the PSF in motion blurred iris images. Because the coarse initial PSF is refined later, the motion contrail can be assumed to be linear for efficiency. First, the inner and outer boundaries of iris [4] are localized. Then, a candidate area is extracted from the center of the pupil. For the simplicity and speed, the directional filter [11] is applied in IDPR to determine the direction coarsely. The length of the specular reflection region in major axis is measured to obtain the length of motion d in E- quation (7). 4. PSF refinement In IDPR, PSF refinement is the main contribution. After initialized, the PSF is refined by the way that L and PSF are optimized alternatively. For efficiency and robustness, we introduce a new way for PSF estimation, a novel scheme

4 to select the gradient maps and a regularization term using initial PSF. In addition, both L and PSF estimation apply an accurate noise model to address the ringing artifacts Selection of gradient maps {L x, L y } are used to represent the selected gradient maps of image L in x and y directions in this paper. Firstly, Equation (5) is solved by using the PSF in last iteration to update L. Then, {L x, L y } are obtained by selecting the most effective gradient region. Finally, the selection of {I x, I y } is accomplished by finding the corresponding regions of {L x, L y }. Selecting {L x, L y } follows three rules. The first one is selecting the effective regions mentioned in [12]. Based on experimental results, Levin et al. [12] contend that not the whole image are appropriate for estimating L but only some regions are effective when sparse prior is used. As described in [12], the effective regions always distribute around the strong edges. The second rule is exclusion of eyelash regions, because it is unavoidable for PSF estimation to be influenced by the less strong edges around the strong ones. For the iris images, the regions of dense and multiple edges are found to only distribute around eyelashes which can be exclude by iris segmentation. Exclusion of overflow regions is the third rule. Once the photons collected by a pixel on camera sensor are overflow, the charge caused by additional photons will have no effect on the pixel value, resulting in the overexposed pixel value. The boundaries of overflow regions are always strong edges that dominate the result in PSF estimation. The lost information of overflow regions leads the PSF estimation to an unsuitable solution. We first apply overflow checking to the current estimated L by using a self-adaptive threshold. Afterwards, the mask of eyelashes region obtained by the method [8] is used to exclude the influence of dense eyelashes. Finally, a gradually decreased threshold is applied to select gradient maps with large magnitude. After these three rules are applied, the selected gradient maps{l x, L y } can be obtained. The selection of {I x, I y } can be accomplished by using the selected gradient maps {L x, L y } and PSF in last iteration. For I i {I x, I y }, we select the regions where the convolution result of L i and f is larger than a threshold PSF estimation PSF estimation calculates f from the selected gradient maps. In most previous methods, only the noise in pixel s value is modeled as a random variable following Gaussian distribution. The spatial randomness of noises cannot be captured totally, so the ringing artifacts always occur. To address this problem, our likelihood term in Equation (4) is based on the model proposed by Shan et al. in [14], which uses image derivatives as a constraint. Although we have the steps of eyelashes and overflow regions exclusion, some regions may not be excluded completely. To be robust, the term f f 0 2 is used, because it can be optimized very fast by pixel-wise operations in the frequency domain. Taking all terms in, the energy function can be obtained as, E f (f) = i ω i i I f i L 2 + α f f 0 2, (8) where i { x, y, xx, yy, xy } denotes the partial derivative operator in different directions and orders, ω i {ω x, ω y, ω xx, ω yy, ω xy } is a weight for each partial derivative and f 0 is the initial PSF L estimation In this step, L is estimated from the given f and I. It is a non-blind deconvolution. For this problem, there are some previous methods which can be applied. Because a sophisticated PSF estimation method is proposed in IDPR, the performance improvement by using a complicated method of L estimation is not significant. In Equation (5), the likelihood also applies the same model as that used in PSF estimation, and the regularization term applies the simple L2-norm constraint. The energy function can be written as, E L (L) = i 5. Experiments ω i i I f i L 2 + β L 2. (9) The proposed method is applicable in both categories of blurred iris images, motion blurred and defocused. So there are no benchmark algorithms comparable with the proposed method in this paper, instead we have to compare the performance metrics of our algorithm with others step by step Experimental setups Ordinal Measures (OM) [15] are used for iris recognition and two performance measurements i.e. equal error rate (EER) and decidability index (d ) are used to quantitatively assess the recognition performance. EER refers to the point in receiver operating characteristic (ROC) when FAR is equal to FRR. d represents the separability of genuine Figure 4. The eye regions from two iris images in our collected database (left) and in ICE 2005 (right) [1].

5 and impostor distributions [13], which is defined as, d = µ pos µ neg, (10) (σpos 2 + σneg)/2 2 where µ pos, σ pos, µ neg and σ neg represent the means and standard deviations of genuine and imposter matching distributions respectively. The larger d value indicates better recognition performance. By applying IrisGuard H100 [2], four datasets are collected by ourselves to conduct the succeeding experiments. The first one contains 266 iris images in different degrees of motion blur, while the second one contains 141 clear iris images captured from the same eyes in the first dataset. For convenience, we call these two datasets MB (motionblurred) and MC (motion-clear) respectively. The third dataset called DB (defocus-blurred), contains 1,160 blurred iris images from 99 eyes of 50 subjects, where most images are defocus blurred. In the last dataset, namely DC (defocus-clear), there are 1,117 clear images from the same eyes in DB. Two iris images used in experiments are shown in Figure 4. In later sections, IDPR-defocus and IDPRmotion are used to represent IDPR only having defocus subunit and motion sub-unit in PSF initialization respectively Experimental results To evaluate the performance of IDPR, we conduct four experiments. The first one is designed to verify the suffering performance caused by blurred images. All of the 1,117 iris images in DC are encoded to perform a total of 5,932 intraclass comparisons and 617,354 interclass comparisons. To evaluate the performance of blurred images, images in DC and DB are used for enrollment and test respectively. There are 13,357 intraclass comparisons and 1,282,363 interclass comparisons. The genuine and imposter matching distributions in terms of the Hamming distance (HD) are computed. Similarly, these operations are implemented on MB and MC. From the results in Table 1, we can see that the recognition performance decreases significantly when the iris images are blurred. In the second experiment, we evaluate the performance of our method dealing with the motion blurred iris images (IDPR-motion), and compare the results with the method in [11]. All the images in MC are used for iris template enrollment. To evaluate the performance of our method and Case1 Case2 Case3 d EER Table 2. Results on MC and MB of motion blur. Case 1: The original images. Case 2: The deblurred images using method in [11]. Case 3: The deblurred images using IDPR-motion. compare it with the method in [11], we take the original images in MB and the deblurred images processed by these two methods for test respectively. The results are shown in Table 2 and Figure 5. It is clear that the proposed method can achieve much more improvement of iris recognition accuracy than [11]. By using the proposed motion deblurring method, the EER of iris recognition drops 60 percents, which is 3 times more than the method in [11]. In the third experiment, the performance of our method on ICE 2005-Left [1] of defocused blur (IDPR-defocus) is assessed and compared with the method proposed in [10]. All the images in this database are encoded to perform a total of 14,653 intra-class comparisons and 1,150,448 inter-class comparisons. Then, the images delurred by our method and by the previous method are used for comparison of improvement of iris recognition performance. The results are shown in Table 3 and Figure 5. We can find that the proposed method outperforms the previous method in both d and EER. The datasets DB and DC are used in the last experiment. Firstly, we verify the effectiveness of IDPR restoring the defocus blurred images again, and the results are corresponding to Case2 and Case3 in Table 4 and Figure 5. Then, expecting for better performance, the whole algorithm proposed in this paper is applied, and the result are corresponding to Case4. From the results in Table 4, three observations can be obtained. First, the proposed method outperforms the previous method when all the images in DB are treated as defocus blurred or clear. The reason is that we refine the PSF to be more accurate in a few iterations. Second, when the whole proposed method is applied, the better performance can be obtained, because the motion blurred images in DB can also be restored effectively. Third, compared with the results showed in previous work [9, 10, 11], we do not get a fantastic improvement on performance. The main reason is that the Ordinal Measures (OM) [15] is adopted to extract the features in this paper, which is more robust to Case1 Case2 Case3 Case4 d EER Table 1. Results with and without blur. Case 1: The clear images (DC). Case 2: The blurred images (DC and DB). Case 3: The clear images (MC). Case 4: The blurred images (MC and MB). Case1 Case2 Case3 d EER Table 3. Results on ICE 2005 [1] of defocus blur. Case 1: The original images. Case 2: The deblurred images using method in [10]. Case 3: The deblurred images using IDPR-defocus.

6 Case1 Case2 Case3 Case4 d EER Table 4. Results on DC and DB. Case 1: The original images. Case 2: The deblurred images using method in [10]. Case 3: The deblurred images using IDPR-defocus. Case 4: The deblurred images using IDPR. blurred iris images than Gabor-based filters. 6. Conclusions In this paper, a novel image deblurring method with PS- F refinement (IDPR) is proposed to enhance the quality of both defocused and motion blurred iris images automatically, so iris recognition accuracy can be significantly improved and the effective capture range can be enlarged. The method is advantageous over state-of-the-art image deblurring methods in the following points. 1) IDPR is more general and can be used for a wide range of applications, s- ince it needs no information of sensors, imaging distance or inter-frame information. 2) An advanced structure consisting of three parts is applied. 3) Using the selected gradient maps and the accurate noise model, the initialized PSF can be refined in an efficient way. 4) IDPR makes full use of the characteristics of iris image for blur estimation, gradient maps selection and regularization term. 5) To our best knowledge, IDPR is the first iris image deblurring method which not just use the prior knowledge or assumption of the blurring form, can deal with both defocus and motion blur. In the future work, we plan to develop a more efficient way to refine the PSF including a better method for deconvolution. In addition, a large database for research and evaluation of iris image deblurring methods will be constructed. Acknowledgements This work is funded by National Natural Science Foundation of China (Grant No , ) and In- Figure 5. ROC curves on MC and MB (left) and on ICE 2005 (right). Case 1: The original images. Case 2: The deblurred images using the previous method. Case 3: The deblurred images using the proposed method. ternational S&T Cooperation Program of China (Grant NO. 2010DFB14110). References [1] ICE. iris.nist.gov/itl/iad/ig/iris.cmf. 4, 5 [2] IrisGuard H , 5 [3] S. Cho and S. Lee. Fast motion deblurring. In ACM Transactions on Graphics (TOG), volume 28, page 145. ACM, [4] J. Daugman. High confidence visual recognition of persons by a test of statistical independence. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 15(11): , [5] J. Daugman. Statistical richness of visual phase information: update on recognizing persons by iris patterns. International Journal of Computer Vision, 45(1):25 38, [6] J. Daugman. How iris recognition works. Circuits and Systems for Video Technology, IEEE Transactions on, 14(1):21 30, [7] R. Fergus, B. Singh, A. Hertzmann, S. Roweis, and W. Freeman. Removing camera shake from a single photograph. ACM Transactions on Graphics (TOG), 25(3): , [8] Z. He, T. Tan, Z. Sun, and X. Qiu. Toward accurate and fast iris segmentation for iris biometrics. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 31(9): , , 4 [9] X. Huang, L. Ren, and R. Yang. Image deblurring for less intrusive iris capture , 5 [10] B. Kang and K. Park. Real-time image restoration for iris recognition systems. Systems, Man, and Cybernetics, Part B: Cybernetics, IEEE Transactions on, 37(6): , , 3, 5, 6 [11] B. Kang and K. Park. Restoration of motion-blurred iris images on mobile iris recognition devices. Optical Engineering, 47:117202, , 3, 5 [12] A. Levin, Y. Weiss, F. Durand, and W. Freeman. Understanding and evaluating blind deconvolution algorithms [13] N. Macmillan and C. Creelman. Detection theory: A user s guide. Lawrence Erlbaum, [14] Q. Shan, J. Jia, and A. Agarwala. High-quality motion deblurring from a single image. In ACM SIGGRAPH 2008 papers, pages ACM, , 2, 4 [15] Z. Sun and T. Tan. Ordinal measures for iris recognition. IEEE transactions on pattern analysis and machine intelligence, pages , , 5 [16] G. Surya and M. Subbarao. Depth from defocus by changing camera aperture: A spatial domain approach. In Computer Vision and Pattern Recognition, Proceedings CVPR 93., 1993 IEEE Computer Society Conference on, pages IEEE, [17] Z. Wei, T. Tan, Z. Sun, and J. Cui. Robust and fast assessment of iris image quality. Advances in Biometrics, pages ,