Biometric authentication using local subspace adaptive histogram equalization

1 Introduction

With the advancement in the technology and boost in biometric applications, lot of research is going on in this field. Now-a-days biometric authentication systems are inbuilt in most of mobile and cellular phones, in laptops, or in services like mobile banking, security systems. In these types of devices, most common biometric modalities are fingerprint, palmprint and face. A typical biometric authentication system involves data acquisition, feature extraction followed by matching and decision making. The main steps carried out in biometric systems are shown in Fig. 1.

Most of these devices have inbuilt camera to capture the picture. The quality of picture depends upon quantity of light, illumination and resolution of capturing device. For proper verification and identification, there must be some image enhancement techniques that would improve the quality of images. The most common method for image equalization is histogram equalization [1]. These are relatively simple methods. It is assumed that quality of image does not vary and has fixed gray-scale mapping which in turn gives same enhancement over the full image. To improve this enhancement technique, the dependency of image quality is shifted to each pixel value and equalization becomes the function of its surrounding gray-scale distribution [2]. Due to their adaptive nature these techniques are known as adaptive histogram equalization (AHE). AHE is generally used in most of latest research in biometrics in [3–5]. Also there are several variants of AHE for further improvements which can be studied in [6–8]. It is always preferred to improve the contrast of the image for better performance of the biometric applications.

Several other enhancement methods that also proves their superiority even for colored and gray images can be studied in [9–12]. In [12], a device and illuminant invariant histogram equalization technique is discussed which does not affect the color of the images. A dynamic histogram equalization for image contrast enhancement is also a variant of HE studied in [13]. Figure 2 shows method of application of AHE in biometric recognition system. In Fig. 1, only data acquisition stage is given. Before feature extraction, several pre-processing is required to enhance the data samples. Figure 2 clearly shows the stage at which AHE can be applied to improve the recognition performance.

To further improve the performance of biometric system, global adaptive histogram equalization technique can be improved locally by varying the effective area with different effective weights. In this paper, a novel adaptive histogram equalization (AHE) variant is proposed i.e. effective area-AHE (EA-AHE) with weights. This method shows the improvement of the stated biometric authentication identification rate in comparison to typical AHE. To validate the proposed algorithm, IITD palmprint database for left and right hand are used in simulations. The block diagram of proposed method is shown in Fig. 3. Finally it is validated through results that proposed technique improves the results effectively. Hence the performance of biometric system is improved considerably.

The paper is organized as follows: Section 2 describes Adaptive histogram equalization. Section 3 presents the Algorithm-effective area-AHE(EA-AHE). Section 4 demonstrates simulations and result analysis. Section 5 presents contribution of the paper. Lastly Section 6 concludes the suggested work.

2 Adaptive histogram equalization

Histogram equalization [1] is a relatively simple method for image enhancement. There are many typical textbook examples in which histogram equalization significantly improves the quality of images with poor lighting [13]. The standard procedure, in this case, is to re-map grayscales of the image so that the resultant histogram approximates that of the uniform distribution. This procedure is based on the assumption that the image quality is uniform over all areas and one unique grayscale mapping provides similar enhancement for all regions of the image. However, when distributions of grayscales change from one region to another, this assumption is not valid. In this case, an adaptive histogram equalization technique can significantly outperform the standard approach [2, 16]. The main idea in adaptive histogram equalization is to find the mapping for each pixel based on its local (neighborhood) grayscale distribution which is a function of the intensity values immediately surrounding the pixel. The number of times that this calculation should be repeated is the same as the number of pixels in the image. This rises to an extensive computation requirement, which even with some modification, cannot be utilized for real-time image enhancement. The method that compromises between global histogram equalization and fully adaptive algorithm is the regional histogram equalization. In this case, the image is divided into a limited number of regions and same histogram equalization technique is applied to pixels in each region.

To apply AHE [17], firstly histogram of a window surrounding the pixel is calculated based on the transforming function which is defined by its cumulative distribution function (CDF). The transformed gray level depends on the surrounding window. If the pixel value is on lower gray level than surrounding pixels then output is maximally black.

3 Algorithm-effective area-AHE (EA-AHE)

The histogram equalization based approaches try to make the distribution of the system uniform by re-assigning the intensity of the pixels. Firstly, the image is normalized to a common range of [0, 1], and p(x) is the density function of intensity distribution of the original image, where x denotes the intensity value of the normalized image. The desired density function of intensity distribution of the output image is equal to 1 after equalization, i.e., AHE formula is AHE = K ∫ 0 x , y p ( u ) du(1)

Steps for applying proposed method:

Divide the image I (x, y) in N subspace windows by superimposing the image with rectangular window function W such as I ¯ ( x , y ) = I ( x , y ) + W. Due to the application of subspace window, the different weights can be applied on different section of image. Thereby, few portions are given more importance in feature extraction.

In experiments, the performance of the proposed method has been evaluated in both verification and identification modes. ROI dimension is set up to 150 × 150 in the experiments. For validation, both left and right hand of IITD palmprint database are used to verify the proposed method. For feature extraction, the orientation of data samples must be same to withdraw the same set of information. There are variations in sample position in database as shown in Fig. 5. To make the hand samples rotation invariant, some pre-processing is required. For this, the coordinates of five fingertips and finger valleys considered as the key points and the centroid from each hand image is extracted. With the help of centroid and fingertips, the image is rotated such that the line joining the tips of the index finger and ring finger becomes horizontal. Then coordinates of finger valleys are used to crop the region of interest (ROI) of each finger of size 150 × 150 [18]. K-nearest neighbor is used for calculating the recognition rate in identification mode. For this cross validation is used from six palmprint samples.

The identification results of the method for left and right palmprint databases have been shown in Tables 1 and 2. It is seen in the Table 1 that with w₁ = 1 which is just the normal AHE case, average recognition rate is 92.4%, 93% and 94.5% with mean, AAD and GMF features respectively. On addition of subspace window w₁ = 0.6, w₂ = 0.4, average recognition rate improves to 92.7%, 94.44% and 94.6% with mean, AAD and GMF features. With the increase in the subspace window, performance is further improved with three and four windows. It is seen that EA-AHE+GMF with w₁ = 0.4, w₂ = 0.3, w₃ = 0.3 also give improvement in recognition rate. Accordingly, it can be seen that EA-AHE+GMF with w₁ = 0.4, w₂ = 0.2, w₃ = 0.2, w₄ = 0.2 weights yields the best results for both databases. In this case, the EA-AHE+GMF with w₁ = 0.4, w₂ = 0.2, w₃ = 0.2, w₄ = 0.2 weights gives recognition rate equal to 96.44% and 94.6% for IITD left and Right databases, respectively. On an average 3 to 5 % result improvement is seen with four windows. However, the proposed technique with further increase in the subspace window degrade the performance of the system due to high de-correlation between the sub-parts of image.

Next, the identification power of the effective area-AHE method have been visually analyzed based on the Cumulative Match Curve (CMC). The recognition rate in CMC curves is plotted against best 10 candidates have been illustrated in Fig. 6 for IITD left hand database and Fig. 7 for IITD right hand database. It is seen in the Figs. 6and 7 that performance of EA-AHE is improved due to effecting weighting. The feature extraction methods with EA-AHE has higher convergence than simple AHE.

For verification, equal error rate (EER) is chosen as a performance measurement quantity where its lower values show the high performance of the system. EER is calculated where FAR is equal to FRR where FAR is False Acceptance Rate which is the acceptance of imposters as genuine and FRR is False Rejection Rate which is rejection of true subjects.

Also, area under the curve of receiver operative characterstics-AUC (ROC) is also used to validate the results. The AUC has an important statistical property: the AUC of a classifier is equivalent to the probability that the classifier will rank a randomly chosen positive instance higher than a randomly chosen negative instance. The maximum value of AUC is 1; higher the value better the performance of the system. The verification results of the proposed method based on the EER and AUC criteria have been summarized in Table 3. From Table 3, it is found that the proposed method offers better results on both the databases. In the terms of AUC, GMF based features on proposed technique with w₁ = 0.4, w₂ = 0.2, w₃ = 0.2, w₄ = 0.2 weights give the best results. All the variants of AHE with EA-AHE+Mean, EA-AHE+AAD and EA-AHE+GMF based features outperform the other techniques. It is seen that eer of AHE+GMF is 1.686 which is reduced to 1.102 on the application of EA-AHE+GMF. Similarly, AUC is improved from 0.8901 to 0.9211 for left hand database and from 0.8888 to 0.9044 for right hand database.

5 Contributions of the paper

6 Conclusion

Current research shows that in biometric authentication systems, image enhancement plays a very important role for most of the low resolution image samples. By locally enhancing the data samples, we can also improve the recognition rate of biometric subsystems. In this paper, a novel adaptive histogram equalization (AHE) variant is proposed i.e. effective area-AHE (EA-AHE). Proposed method improved the stated biometric authentication identification rate in comparison to the typical AHE. The proposed method shows the modest improvement in the results. Here global adaptive histogram equalization is improved to a local AHE technique by varying the effective area with different weights. These local AHE methods can also be used in biometric fusion system, thereby increasing the performance of such subsystems and hence can provide more secure applications. To validate the proposed algorithm, IITD palmprint database for left and right hand is used in simulations. Experimental results demonstrate that the proposed method can enhance the results effectively.