Dorsal hand vein biometric recognition: A comparative analysis of feature extraction methods

This study examines feature extraction methods in biometric recognition systems, focusing on their impact on accuracy and performance. Nine methods (CNN, VGG16, ResNet50, LBP, Gabor, GLCM, DCT, FFT, and DWT) are analyzed using the FYO biometric database’s dorsal trait dataset (320 samples). The methods vary significantly in feature counts and extraction times: VGG16 (4,096 features, 75.15s), ResNet50 (2,048 features, 77.82s), and lighter methods like LBP (59 features, 0.82s), DCT (64 features, 0.15s), and DWT (256 features, 0.57s). Frequency-domain techniques show distinct patterns. A key finding is the tradeoff between feature dimensionality and computational cost, guiding technique selection. FFT showed the best balance of high inter-class variance and low intra-class variance, followed by DCT and LBP. CNN remains effective due to its positive mean and low SD. VGG16 and ResNet50 were unstable; Gabor lacked discriminative strength. FFT and CNN represent optimal frequency-based and neural network-based approaches. This work contributes to the understanding of feature extraction methods in biometric recognition by providing a comparative statistical analysis—based on feature dimensionality, extraction time, mean, standard deviation, and coefficient of variation—that can inform the selection of appropriate feature extraction techniques for dorsal hand vein biometric applications.