Intrinsic property-based soft biometric classification using wavelet CNNs and LRP
Sunusi Bala Abdullahi, Kosin Chamnongthai · 2025
The rapid evolution of optical sensors has significantly impacted the field of soft biometrics, necessitating the development of adaptive automatic deep learning methods. Traditional deep-learning approaches for feature extraction and classification often struggle with biases related to sensor position and activity configuration. Recent advancements in deep learning have leveraged attribute representation-based transfer learning for effective biometrics identification. However, these methods are limited in identifying the most contributive features for classification tasks. This study proposes a novel approach to transform human activity features using the effective wavelet transform (eW ), which minimizes sensor biases by converting activity features into a novel representation of intrinsic biometric properties. These features are then learned through a well-trained Wavelet Convolutional Neural Network (eWCNN ). To interpret the eWCNN, we employ Layer-wise Relevance Propagation (LRP) to analyze the features and plot relevance scores. Our method outperforms existing state-of-the-art techniques, providing an interpretable solution for soft biometrics identification and classification.