Enhanced Local Line Binary Patterns for Palmprint Images

In reference [17], an enhanced local LBP (ELLBP) was introduced. This method improves the performance of LLBP; however, it focuses solely on vertical and horizontal line features. In reference [18], an innovative feature extraction method called the Surrounded Patterns Code (SPC) was employed. This approach effectively considers the surrounding patterns of the main features.

In reference [19], the author proposed a Controlled Tiger Optimization (ENcTO) approach using Convolutional Neural Networks (CNNs) for image identification. They applied the Entropy Controlled Tiger Optimization and a multi-scale deep neural network identification method to recognize palmprints and human palms. In the processing flow, the author used feature extraction and finger recognition with a multi-scale deep CNN identifier. The results of this research demonstrate that the ENcTO approach outperforms the state-of-the-art method for gesture identification, achieving a recognition rate of 96.72%.

In reference [20], the author applied Machine Learning with content-based feature extraction methods for optimizing file cluster types of identification. The author used three content-based feature extraction methods, which are the Rate of Change, Entropy, and Byte Frequency Distribution, to produce an image cluster histogram. The ELM classifier was used to evaluate entropy, ROC, and BFD methods for an image file or a non-image file type. The results demonstrate that the proposed approach using a combination of the 3 feature methods produces high identification accuracy (93.46%) in classifying the file type.

In reference [21], the author proposed a support vector machine (SVM) identification method to identify image and non-image file clusters. The SVM identifies applied based on three entropies, byte frequency distribution, and the rate of change content-based feature extraction method. Radial basis and polynomial kernel functions are used to optimize the identification of image cluster content. The results show that the accuracy (96.21%) of identification of the SVM identifier by the polynomial function, and the accuracy (57.58%) of the SVM classifier by radial basis function.

In reference [22], the authors propose a new approach using the LBP method and the Gabor content-based feature. Also, using K nearest neighbor identifiers used with these features of a finger vein. Three datasets were used for the experiment. The results show that the proposed method outperforms compared with the state-of-the-art approaches. The proposed approach achieved the lowest identification rate (0.039%) using the SDUMLA, the lowest identification rate (0.064%) using the FV-USM, and the lowest identification rate (0.037%) using the MMCBNU_6000, without using any enhancements.

In reference [23], the author presented a new approach using the LBP operator for image face recognition. The LBP is used by distributing the face image into non-overlapped regions. The results show that the proposed approach has a 97.5% rate using Olivetti Research Laboratory databases.

In reference [24], a new approach based on the LBP method and the support vector machine method is presented. The results show that the LBP demonstrates plant classification accuracy with 91.85% using four subclasses, background, canola, corn, and radish, with a 24,000-image dataset.

In reference [25], the author proposed an automatic approach to detect the suspicious regions of interest for breast dynamic contrast-enhanced (DCE) MRI based on region growing. The results show that the new approach is able to identify suspicious regions on the PIDER breast MRI dataset.

In reference [26], the author proposed the VisGIN approach using and passes Visibility Graphs (VG). GINConv is used for the convolutional layers. The Visibility Graph (VG) is used for input. The results show that the fruition of the approach achieves 99.76% in the grant-access decisions evaluation. Therefore, the authors' VisGIN approach demonstrates the effectiveness for time-series binary identification tasks by using graph machine learning methods.

In reference [27], the authors proposed a novel method named the VGG16-PCA-NN approach. The main objective of the proposed approach is to improve the precision of facial authentication. To extract features, the author used the VGG16 model as a pre-trained model on the ImageNet dataset. The method improves reliability under varying conditions by handling environmental and physical challenges. Its high accuracy across multiple modalities supports secure, real-time biometric identification.

This paper introduced a new feature extraction method called MLLB. This approach can analyze the texture features of the palmprint by handling multiple directions of palmprint lines. Specifically, it first examines the vertical and horizontal directions using the equations outlined in reference [17].

$\begin{aligned} L L B P_H= & \sum_{n=1}^{c-1} s\left(H_n-H_c\right) 2^{(c-n-1)} +\sum_{n=c+1}^N s\left(H_n-H_c\right) 2^{(n-c-1)}\end{aligned}$            (1)

$\begin{aligned} L L B P_V= & \sum_{n=1}^{c-1} s\left(V_n-V_c\right) 2^{(c-n-1)} +\sum_{n=c+1}^N s\left(V_n-V_c\right) 2^{(n-c-1)}\end{aligned}$            (2)

where, Hn is the number of pixels in a horizontal vector, Hc represents the horizontal vector center pixel, c indicates the position of the center pixel, and N is the length of the vector in pixels. This is a value of a vertical direction, Vn is the pixels number in a vertical vector, Vc is the center pixel of a vertical vector, and can be defined as follows: V_n is the number of pixels in a vertical vector, V_c is the vertical vector center pixel, and s(x) can be defined as follows:

$s(x)= \begin{cases}1, & x \geq 0 \\ 0, & x<0\end{cases}$            (3)

Secondly, the proposed MLLBP method considers the primary and secondary diagonals according to the following Eqs. (4) and (5):

$\begin{aligned} L L B P_{D 1}= & \sum_{n=1}^{c-1} s\left(D 1_n-D 1_c\right) 2^{(c-n-1)} +\sum_{n=c+1}^N s\left(D 1_n-D 1_c\right) 2^{(n-c-1)}\end{aligned}$            (4)

$\begin{aligned} L L B P_{D 2}= & \sum_{n=1}^{c-1} s\left(D 2_n-D 2_c\right) 2^{(c-n-1)} \quad+\sum_{n=c+1}^N s\left(D 2_n-D 2_c\right) 2^{(n-c-1)}\end{aligned}$              (5)

where, $L L B P_{D 1}$ is a value of a primary diagonal direction, D1_n is pixels of a primary diagonal vector, D1_c represents the center pixel of a primary diagonal vector, $L L B P_{D 2}$ is the value of a secondary diagonal direction, D2_n is pixels of a secondary diagonal vector, and D2_c represents the center pixel of a secondary diagonal vector.

Consequently, combinations are applied between the horizontal and vertical values, and between the primary and secondary diagonal values. The combination is implemented according to the following calculations:

$E L L B P_1=v_1 \times L L B P_H+v_2 \times L L B P_V$         (6)

$E L L B P_2=v_1 \times L L B P_{D 1}+v_2 \times L L B P_{D 2}$            (7)

where, $E L L B P_1$ represents the combination of the horizontal and vertical values, $E L L B P_2$ represents the combination of the primary diagonal and secondary diagonal values, and $v_1$ and $v_2$ are weighted summation parameters [19].

Hence, to attain the effective feature extraction value from the $E L L B P_1$ and $E L L B P_2$, competitive method is considered between them.

Empirically, the following equation can be useful to compute:

$I L L B P_T=\min \left(E L L B P_1, E L L B P_2\right)$         (8)

where, $I L L B P_T$ represents the effective feature extraction value, and min represents a minimum operation selection.

An example of the main MLLBP processes is given in Figure 4. All the above calculations are used in each single pixel of the palmprint image. Afterwards, the featured palmprint images are collected. To complete collecting the data variance information of the 0 palmprint image, the following equations are considered [20] after partitioning each featured image into multiple blocks:

$M_{b l}=\frac{1}{j}$             (9)

$S T D_{b l}=\sqrt{\frac{1}{j-1} \sum_{i=1}^j\left(b l_i-M_{b l}\right)^2}$           (10)

$C O V_{b l}=\frac{S T D_{b l}}{M_{b l}}$              (11)

where, $j$ represents several pixels, $b l$ represents a pixel ( $5 \times 5$ pixels as in references $[21,22]$ ), $M_{b l}$ represents an average value of each block of pixels, $S T D_{b l}$ is the standard deviation value of each pixel and $C O V_{b l}$ is an achieved Coefficient of Variance ( COV ) value. COV values have many benefits.

4a.png

4b.png

Figure 4. An example of the main MLLBP processes

To illustrate, it can efficiently provide variances among pixel values; its computations are simple; it can decrease the size of the image [17]; all computed values are positive, and the variances among image features are efficiently defined for different subjects [26-34]. So, a vector of COV values can be collected for each palmprint image. These values are used as inputs to a Probabilistic Neural Network (PNN). The main architecture of PNN is shown in Figure 4.

It consists of input, hidden, summation, and decision layers. Also, this network, like other neural networks, works in two phases: training and testing.

Important benefits of using a PNN can be observed [27]: it is so fast during the train; it does not need more than one epoch; it is not suffer from the local minima problem as in the Multi-Layer Perceptron (MLP) neural network; its architecture is so flexible, where it can easily remove or add training data; and the input training information are stored in training weights.

In this article, we introduce the MLLBP as a novel feature extraction method for contactless palm print recognition. By expanding traditional Local Binary Patterns to analyze not only horizontal and vertical directions but also primary and secondary diagonals, MLLBP captures more detailed textural information. Our competitive matching strategy, which selects the minimum response among horizontal, vertical, diagonal, and antidiagonal analyses, has proven more effective at highlighting the most discriminative linear features. An experimental evaluation on a 3D/2D PolyU hand image database demonstrated the effectiveness of MLLBP. Using coefficient of variation (COV) feature vectors fed into a probabilistic neural network, we achieved an equal error rate (EER) of just 0.40%. This result not only exceeds previous linear methods such as LLBP and ELLBP (each with EERs of 0.80%) but also significantly outperforms patch, histogram, and surrounding pattern approaches, which have EERs ranging from 1.40% to 6.40%. Besides offering superior verification accuracy, MLLBP shows robustness to changes in illumination and low image resolution, making it especially suitable for practical and cost-effective biometric systems. The simple block-based COV computation ensures compact feature representation and quick processing, while the PNN classifier allows fast training and avoids issues with local minima.

In the future, we plan to extend MLLBP to multispectral or three-dimensional palm print data and integrate it with deep learning frameworks for end-to-end feature learning. Additionally, large-scale testing on more diverse populations and real-world capture conditions will help verify the method's generalizability and feasibility for deployment. Additionally, we aim to extend the proposed method to multispectral or three-dimensional palm print data and incorporate it into deep learning frameworks for end-to-end feature extraction. Specifically, Convolutional Neural Networks (CNNs) can be used to automatically learn hierarchical spatial features from palm print textures. Meanwhile, the Siamese CNN architecture might be effective for one-shot learning in verification tasks.