ECG Biometric Recognition Based On Feature Learning And Multi-feature Fusion

Biometric recognition refers to identity technology that uses human physio-logical features such as faces,fingerprints,iris and veins,or behavior features such as gait,voice and handwriting.Compared with traditional identity recog-nition such as passwords and cards,biometric recognition has many advantages such as simplicity and quick speed,never forgotten,safe and reliable.Biometric recognition has great social and economic benefits with wide application future,and the academia and industry are very interested in it.Electrocardiogram(ECG)biometric recognition is a new identification technology based on human biological features.Compared with other biological features,ECG signals as a biological feature inside the human body has unique advantages such as the live-ness detection,high security,small data size and hybrid information.In recent years,with the emergence of portable and mobile ECG signal acquisition devices with low power consumption,small size and no conductive glue,ECG biomet-ric recognition has attracted widespread attention and is in the hot spot of the current research.Due to the influence of many factors such as the signal acquisition devices,body position and acquisition environment,the ECG signals often contain many noise.The sources of ECG signal noise are multiple and their effects are compound,and the existing de-noising methods do not meet demand.In addition,the ECG signals are easily influenced by physical and psychological changes,which can lead to stretch or contract,and the waveform of heartbeats from the same person may vary over time,which is called intra-class variation.The ECG signal has the problems of noise and intra-class variation,which have brought obstacles to the practical application of ECG biometric recognition,and new technical means are urgently needed to solve it.This thesis studies the effective methods to solve the problems of ECG noise and intra-class variation from the perspective of feature learning and multi-feature fusion,and improves the robustness and effectiveness of ECG biometric recogni-tion.The main works and contributions of the thesis are as follows:1.To address the issue that the traditional sparse representation learning based ECG recognition methods are not quite robust to noise and intra-class variations in the case of small samples,we propose a multi-scale deep cascade bi-forest model for ECG biometric recognition.The proposed method includes two parts:multi-scale signal coding and deep cascade coding.In the former,we design the multi-scale feature extraction and adaptive weighted pooling operation,which can improve the ability of characterization to reduce noise impact.In deep cascade coding,we extract the complementary information of different levels to obtain the feature with higher semantics and discrimination,which helps to reduce the impact of intra-class variations on the performance of ECG biometric recognition.2.To address the issue that the existing multi-view subspace learning methods are sensitive to noise,we present a multi-view discriminant analysis approach in the consideration of sample diversity for ECG biometric recognition.Based on the multi-view discriminant analysis,the proposed method obtains the discriminative information of samples diverse by constructing intra-class compact and inter-class discrete graph,which can improve the discriminant analysis.To eliminate redundant information,we introduce a de-noising constraint to enhance further the learning performance of multiple views3.To address the issue that the learning performance of multi-feature joint sparse representation is degraded with noise and intra-class variation,we propose a unified sparse representation framework which collaboratively exploits joint and specific patterns for ECG biometric recognition.The proposed method con-structs the objective function based on multiple constraints.It not only considers the consistency and pairwise constraints of multi-feature sparse representations to improve the recognition performance,but also combines special constraints different features,which is helpful to eliminate noise and intra-class variations.4.To address the issue that the existing methods of multi-feature collec-tive nonnegative matrix factorization rarely consider the noise of the semantic space,we propose a robust multi-feature collective nonnegative matrix factoriza-tion model to handle noise and sample variation in ECG Biometric recognition.To enhance the discrimination of learned representations,we integrate label in-formation and multiple norms in the proposed model,which not only preserves intra-and inter-subject similarities but also mitigates the influence of noise and intra-class variation.