A Finger Vein Recognition Framework Using Foreground–Background Decomposition and Translation-Invariant Encoding

Abstract

While deep learning–based layered feature extraction methods have achieved remarkable success, their reliance on large-scale annotated datasets limits their applicability in small-sample scenarios. To address this challenge, a novel feature extraction method has been proposed within the traditional image processing framework. This technique is specifically designed for scenarios with limited training data, aiming to enhance performance and efficiency in such conditions. Inspired by image separation algorithms and multifeature fusion strategies, the proposed approach employs guided filtering combined with the Sobel gradient operator to decompose the original finger vein image into a foreground layer and a background layer. Texture features are extracted from the foreground layer, while structural features are derived from the background layer, resulting in two complementary feature maps that capture multidimensional information. These maps are then encoded into a unified one-dimensional feature vector using block-wise histogram descriptors, which enhances feature representation and ensures translation invariance. By separately extracting and effectively fusing multilevel features, the method significantly alleviates the impact of noise on feature extraction and discriminative performance. Without relying on large-scale data, it improves the robustness and practicality of finger vein recognition. Extensive experiments on public datasets validate the effectiveness and generalization capability of the proposed approach.