Histogram of Oriented Gradients Feature Extraction in Enhanced Convolutional Neural Networks Method for Palm Vein Recognition

Histogram of Oriented Gradients, or HOG, is an efficient method for extracting features from images that shows to be reliable in recognizing and detecting objects in a variety of conditions. HOG is a very useful tool in image analysis because it is invariant to brightness, rotation, and scale transformations and can generate detailed local descriptors. By exploiting the advantages of HOG in feature extraction, the palm vein recognition system can overcome the weaknesses of palm vein images, which are sensitive to changes in brightness and rotation. By analyzing the distribution of orientation gradients, HOG can extract important features from palm vein images generated from infrared light, enabling reliable biometric recognition. We proposed a method applying HOG into palm vein biometric system with Convolutional Neural Network (CNN) which has strong hierarchical representation capabilities, automatic feature extraction, and the ability to handle high-dimensional data. Our proposed model is tested on three public contactless palm vein datasets (CASIA and VERA). We evaluated the results using various performance biometric measures, such as correct recognition rate (CRR) accuracy, and area under curve (AUC) of the receiver operating characteristic. The simulation results showed that HOG features could improve accuracy from pretrained CNN such as AlexNet, VGG-16, ResNet-50 and EfficientNet.

I. Introduction

The palm vein is a popular biometric to research because of its location and convenience during the feature collection process. The advantage of the palm vein feature is that it is difficult to duplicate because the skin of the palm covers it, leaves no trace on an identification device, and is resistant to changes in weather, temperature, and age. Each human has palm veins different from anyone else's, even twins [1], so the palm vein can be prospectively developed for personal recognition. There are several stages in the palm vein biometric system, including image acquisition, preprocessing, feature extraction, and classification. The image acquisition should be completed by using a special camera because the palm vein is visible to be captured in infrared waves, especially between 750nm -1mm [2]. Infrared waves can penetrate skin layer and be absorbed by deoxidized hemoglobin in the blood vessels [3]. The camera captures an image with the original palm vein image quality, which must be improved due to the low contrast and non-uniform illumination. Thus, the image requires image preprocessing, which involves a binary image, segmentation, filtering, and contrast enhancement. After preparing the palm vein image, feature extraction and classification are commenced. The process of recognizing a person's palm veins can be done with learning, several methods of which support the palm vein recognition process, including artificial neural networks (ANNs). However, the development of research on ANNs to recognize objects has limitations in feature extraction [4]. Currently, deep learning is popular for biometrics [5], [6], and it has been widely applied in such biometric fields such as speech recognition [7] [8], handwriting [9], [10], face recognition [11], as well as fingerprint [12], iris [13], and vein biometrics [14].