Multispectral palmprint recognition using Pascal coefficients-based LBP and PHOG descriptors with random sampling

Palmprint is a relatively newly discovered physiological biometric trait that has recently arisen as an active area of study. The rich features of the palmprint are the key to its recognition power. Patterns elicited from palms have excellent discriminatory power as they have more features on the surface than fingerprints while being stable. However, it should be noted that observations of palmprint features are often affected by various issues, i.e. variations in lighting, orientation and noisy sensors, which make the task of identification more complex. Variance in illuminations in particular can seriously affect the ability of systems to recognize individuals. The majority of palm recognition methods can be regarded as sufficiently robust to deal with all variations in image conditions. However, researchers today continue to attempt to solve this problem and develop systems that can be used accurately to identify a person.

In the literature, a number of techniques are reported for use in palmprint recognition, classified into several different categories: structure based, statistic based, subspace based and code based. Structure-based algorithms mainly concern information on the direction and location of the main lines and folds in the palmprint, such as principal lines, wrinkles, delta points and minutiae. Structure-based algorithms are the traditional approaches to extract the features of palmprints and provide effective representation and matching. For feature extraction, there are many proposed approaches which use a variety of line detection operators. Funada et al. [1] presented an algorithm that extracts high probability local palmprint features, such as ridges, by eliminating the creases. Zhang and Shu [2] attempted to determine datum points from the main lines using a directional projection technique. These datum points are found to be rotation and translation invariant due to the steadiness of the main lines. Sobel and morphological operations were used in [3] to extract line features from palmprints. For the representation of the features, this method primarily uses straight line segments or feature points instead of ridges. Housdorrf and Euclidean distances are widely used for the matching process.

In the case of statistic-based palmprint identification, the works that have been published include local or global statistical approaches. Systems based on the local features of the palmprint include discrete cosine transform, Fourier transform, wavelet transform and Gabor transform. These tools have been studied and used to transform images before feature extraction task in order to extract more distinctive features. In [4], the authors transformed a palmprint image into the wavelet domain and computed the average and variance of each patch to create a normalized palmprint vector. The standard deviation of the small block is used as a feature. In [5], the mean and standard deviation of the small patch are employed as a feature after transformation with a Gabor filter. In [6], a histogram of a local binary palm image was used as a palmprint feature. Global statistical approaches compute the global features of the palmprint, such as moments, centres of density and gravity , directly on the palmprint image. Correlation coefficients, first-order norms and the Euclidean distance are often used for the purpose of matching. A palmprint recognition system is proposed in [7] using Hu invariant moments as patterns on an Otsu binarized palmprint.

Other efforts in this domain have also explored a variety of subspace-based algorithms to derive a compact feature subspace for palmprint data. The main subspace approaches reported in the literature employ a palmprint as a high-dimensional matrix and mapping it to a lower-dimensional matrix. Then, the generated images can be represented and matched in this low-dimensional space. Subspace-based approaches include linear and nonlinear space algorithms. Lu et al. [8] introduced PCA effectively in palmprint recognition. Notwithstanding the significant achievements of PCA, some challenges remain requiring more investigation. In [9], two-dimensional PCA (2DPCA) was successfully introduced for palmprint recognition. This method relies on a two-dimensional palm image matrix rather than one-dimensional vector, and a palm covariance matrix is generated directly employing the original palm matrices. In [10], Niyogi suggested locality preserving projection (LPP). The aim of LPP is to solve a generalized eigenvalue problem. It seems to be more stable to noise than PCA and LDA [10]. Researchers have also proposed a number of approaches based on coding to extract features for palmprint recognition. These include Fourier transform, the Gaussian derivative filter, wavelet transform and Gabor wavelet transform. Among these methods, Kong and Zhang [11] suggested a fusion code algorithm to encode the Gabor filter phase using six directions. Moreover, based on ordinal code, Song et al. [12] proposed a phase coding scheme using 2D orthogonal Gabor filters. These are employed for various directions to extract texture features, and a phase coding algorithm is applied to describe the palm image. Another approach, discussed in [13], introduced a robust line orientation code (RLOC) for palmprint recognition as an improved version of the competitive code. In the proposed approach, the LBP technique is customized based on conventional thresholding using Pascal’s coefficients of order n [14]. The proposed variant called Pascal coefficient LBP (PCLBP) descriptor is inspired by the SLBP descriptor [15, 16] . This allows us to detect only the robust patterns from the palmprint images. This approach has many advantages, such as the simplification of implementation and high-speed computation. The main idea is to use a varied number of intervals to generate a distribution of binary codes for every pixel position thus creating more robust descriptors to cope with the changing image distortions. In the proposed variant, the main difference from LBP is that the threshold value is tuned using Pascal’s coefficients of order n with an alternating sign. Furthermore, this variant is also extended to MLBP in this paper, referred to as the Pascal coefficient MLBP (PCMLBP) descriptor, for which the PCMLBP features of the different scales are first extracted and their histograms subsequently concatenated into a long feature. Furthermore, to achieve higher recognition rates, we propose a novel feature method to form a new set of features based on the combination of the pyramid histogram orientation gradient (PHOG) descriptor with the PCLBP descriptor, so that the histogram bins have a more powerful discriminatory capability. Nevertheless, having a large number of features can become a curse in terms of classification. To solve this problem, PCA is used to reduce the size of the dimension of the vector of palm features. In addition, we construct a multiple LDA classifier from many individual clarifiers. A powerful decision rule is used for the purposes of combination and is known as ensemble learning. LDA can be achieved by maximizing the ratio of the determinant of the within-class variance and the determinant of the between-class variance. The assessment of the performance of our proposed approaches was conducted using the multispectral palmprint database available from the Hong Kong Polytechnic University (PolyU), using LDA classification. In addition, a comparative analysis of our proposed algorithms was performed against a number of state-of-the-art counterparts, including the techniques described in [17‐20]. The main body of this paper is composed of six main sections. Sections 2 and 3 describe the different steps constituting our proposed multispectral palmprint recognition approach, whereas the experimental results related to the proposed methods are reported in Sect. 4. Section 5 discusses the computational complexity of the proposed methods. The paper ends with a conclusion and proposals for future work in Sect. 6.

Palmprint recognition can be effectively performed by using the best spectrum wavelength range (blue, green, red and NIR). A typical scheme for such a system comprises of the following steps: palmprint image acquisition, feature extraction, dimensionality reduction and classification. Figure 1 shows the main steps of our proposed method. The palmprint and textures are extracted using PCMLBP and PCLBP–PHOG descriptors. The classification of this new set of features is undertaken using PCA and random subspace LDA. The correct identification is achieved when the test palmprint label matches the palmprint label of the same subject in the training samples. The process may result in misclassification when the test palmprint is classified using a palmprint label of a different subject from the training sample. LDA-based classification results are combined at the fusion stage to obtain higher recognition rates. In this work, an ensemble framework is proposed based on the PHOG and the customized LBP descriptor for the classification of palmprint images. To enhance the accuracy of existing LBP descriptors, Pascal coefficients’ approach is used for the purpose of customization.

The previous section has demonstrated that LDA outperforms common subspace methods. However, there is still plenty of room to improve its performance further. One likely avenue for improvement has been proposed which enhances performance by employing numerous LDAs trained on various parts of the data which are then combined in an ensemble [28].

An ensemble is a method used to merge a number of weak classifiers to build up a strong classifier, as shown in Fig. 3. Such multiple classifier ensembles are variously called mixtures of experts or combinations of multiple classifiers. The idea is to use a set of weak classifiers and assemble them to build an ensemble classifier with a powerful decision rule. It should be noted that weak classifiers do not always make the same error. The overall error of the combined classifiers together provides an outcome that improves considerably on what any individual classifier could achieve on its own. One main condition for the success of combining different weak classifiers is that the outcomes of single classifiers for the same inputs must vary. The diverse individual classifiers are obtained by employing various training datasets, various feature datasets, different types of single classification methods and a fusion rule. A multiple LDA classifier is constructed in this study by combining many individual weak classifiers, aiming to build a more robust LDA classifier that includes the overall palm feature space without loss of discriminant information [28].

Extensive experiments were conducted to validate and evaluate the effectiveness of the proposed methods, using the multispectral palmprint database. The multispectral palmprint database was obtained from the Hong Kong Polytechnic University (PolyU) [29].

The first method used different radii of the PCLBP operator. It is noticeable that there are too many patterns in the multiscale PCLBP, with high computational complexity. The second method is a combination of PHOG and PCLBP, which leads to the dimensionality problem and further increases computational complexity. Thus, the idea was to reduce the dimension size, which has a large number of palmprint features for use in the classification. The dimension of the PCLBP–PHOG feature vector can be reduced by using feature selection or extraction methods. PCA was applied to a randomly selected feature subset in order to reduce both the dimensionality of the palmprint feature vector and the computational complexity involved. The performance of the proposed PCLBP–PHOG method was compared to other methods. The experimental results obtained showed that it gives outstanding results when compared to other existing methods.

A new multispectral palmprint recognition methodology has been proposed based on the combination of PCLBP and PHOG descriptors. The study used random sampling to construct many individual LDA clarifiers and applied PCA for to reduce the dimensions on random length feature vectors before using an LDA classifier for final classification. In this work, we use the fusion rules to fuse a number of LDA classifiers and thus demonstrate the concept of the framework proposed. The experimentation is performed according to two different protocols, confirming that the proposed methodology achieves higher recognition rates. Furthermore, it can be observed that in the PCMLBP approach, both with three and six training samples, there is better performance for the blue spectral band than other spectral bands. Moreover, the PCLBP–PHOG approach achieves a higher recognition rate and outperforms the PCMLBP and other state-of-the-art techniques, with the red spectral band performing better compared to other palmprint spectra. The proposed method, using either protocol, has many advantages, such as being simple to implement. Currently, we are extending this work to finger knuckle print data for validation purposes and to verify the effectiveness of the proposed approach.