Biometric recognition systems have evolved from uni-modal systems to multi-modal frameworks, improving both accuracy and robustness. Deep learning has been at the forefront of this evolution, and Convolutional Neural Networks CNNs have been the foundation for biometric fusion systems due to their ability to tap the spatial features of the input data. However, CNN architectures have inherent challenges in handling long-distance dependencies, as well as inter-modal dependencies, within the biometric modalities. In this narrative review, the architectural development of deep learning models used in multimodal biometric fusion systems will be critically discussed, from the use of CNN-based models and Recurrent Neural Networks RNNs, including Long Short-Term Memory LSTM architectures, towards the development of hybrid and transformer models. This article will also discuss the different levels of biometric fusion, including fusion at the sensor, feature, score, and decision levels, and will synthesize the challenges associated with multimodal biometric fusion systems. Through the analysis of research gaps in existing studies, as well as the motivations behind the transition to new architectures, this literature review points to the combination of CNNs and Transformers as an area of great promise for the development of scalable, transparent, and robust multimodal biometric fusion systems.
Alazawi et al. (Wed,) studied this question.