Research On Multimodal Biometric Fusion Retrieval Algorithms

Human biometric traits have evolved into a potent instrument for public security authorities to identify all types of criminal activity due to their benefits of uniqueness and specificity.High quality data may be supplied to unimodal biometric recognition technology to provide excellent accuracy.Nevertheless,most of the biometric information gathered in criminal investigation cases is of poor quality,and conventional unimodal biometric identification technology seldom suffices to fulfil the demands of these cases.This study presents a multimodal biometric fusion retrieval network model based on face,fingerprint,and iris based on deep learning methods.This paper also develops a multimodal biometric fusion retrieval platform based on the development of virtual homologous multimodal dataset.Specifically focused study includes:First,this paper proposes baseline methods for unimodal biometric retrieval and constructs a virtual homologous multimodal biometric dataset using publicly accessible data.This research creates a virtual homologous multimodal dataset based on a public dataset that contains 2,712 people and 4,0482 photos to overcome the issue of the short size of the public multimodal biometric dataset.In parallel,a sizable real-world dataset is created based on public security data to evaluate the effectiveness of the fusion method in the score level.This research suggests a benchmark technique for unimodal biometric retrieval based on neural networks to more effectively compare the performance of the fusion algorithms.The results of the experiments demonstrate how well the deep neural network-based approaches for retrieving biometric features work.Secondly,a deep learning approach based on pixel level and feature level is suggested for multimodal biometric fusion retrieval.Channel-based fusion,space-based fusion,and intensitybased fusion are proposed for pixel level fusion;joint feature representation methods,methods based on modality-specific low-rank factors and attention mechanisms are proposed for feature level fusion to address the issue of insufficient accuracy of unimodal algorithms for low-quality biometric retrieval in public security combat applications.The results of the experiments demonstrate how well the proposed fusion approaches can make use of the complementary information across the different modalities and enhance the capacity to retrieve biometric features.Lastly,a multimodal biometric retrieval technique based on score level is presented to solve the issues that heterogeneous score unification and realistic score matching are tough tasks for multimodal algorithms.First,the ranking partition collision(RPC)theory is suggested,and the formal formulation of the problem as well as the mathematical viability are derived.Experimental results demonstrate that the proposed method is significantly more efficient and can produce a more valuable ranked list of candidates,which aids investigative technicians in increasing their efficiency and accuracy.The development of a demonstration platform for multimodal biometric fusion retrieval comes in fourth.A visual demonstration platform is designed to realize the functions of modal data input and model inference.This is executed in response to the issue of the unclear value of landing multimodal biometric fusion algorithms in real-world public security practice.The multithreaded feature is used for the pre-extraction of the underlying library features to increase the operation efficiency.According to experimental findings,while GPU acceleration mode is employed,the identification performance of the platform is greatly increased when GPU acceleration mode is employed,with all model inference finished in less than 10 seconds.