Keywords

1 Introduction

Nowadays, people demand for more secured systems and security has become a prime factor [1–14]. Unimodal systems are found to be not very efficient to recognize under uncontrolled environment. This has raised the need for more secure system using multimodal biometrics. Multimodal fusion techniques can be broadly classified into three categories: Score level fusion [15], decision level fusion [16], and feature level fusion [17]. Section 2 describes multimodal fusion techniques, Sect. 3 contains the Result and Discussion and Sect. 4 draws the Conclusion.

2 Multimodal Fusion Techniques

Multimodal fusion techniques are usually classified as: score level fusion [15], decision level fusion [16], and feature level fusion [17].

2.1 Score Level Fusion

Authors in [15] describe taking advantage of the uncertainty concept of the Dempster-Shafer theory; unified framework for multimodal biometric fusion is developed by improving the performance of multibiometric authentication systems. Uncertainty factors affect the recognition performance of the biometric systems. Modeling uncertainty helps to address the confidence of the fusion outcome and uncertainty of data. To improve the fusion a combination of classifier performance and quality measures is proposed to encode the uncertainty concept. Quality measures contribute unequally to recognize performance. Hence, only significant factors are fused with the Dempster-Shafer approach to generate an overall quality. In the success of uncertainty, modeling score plays an important role. In this approach multiple biometric modalities can be effectively fused, and the approach is robust to variations in classifier accuracy and quality, and enables multimodal biometric systems to operate in less constrained conditions [15]. The authors in [15] claim that their proposed approach can effectively fuse multiple biometric modalities, hence it is robust to variations in quality and classifier accuracy, and can enable multibiometric systems to operate in less constrained conditions.

2.2 Decision Level Fusion

Authors in [16] describe decision level fusion for the multimodal biometric system using social network analysis (SNA). Problems like classifier selection, dimensionality reduction, and aggregated decision making can be sought out by employing the decision fusion using SNA. Based on the similarity and correlation of features, among the classes, social networks are constructed. Fisher Linear Discriminant Analysis is used by the authors in [16] as feature extractors to reduce the dimension and to identify significant features. Based on the two levels of decision fusion methods final classification result is generated. When SNA is employed, it reduces the false acceptance rate (FAR) for both single biometric traits and multimodal biometrics. Each decision is made after the improvement of the classifier confidence in the decision fusion scheme [16]. Authors in [16] claim that the method reduces the FARs for both single and multimodal biometric traits when the SNA is employed. In case of the decision fusion scheme, each decision is made after the improvement of the classifier confidence.

2.3 Feature Level Fusion

Authors in [17] consider two biometric traits, i.e., finger-knuckle and finger-nail obtained by the single scan of dorsum hand. In this approach, a combination of finger-knuckle and finger-nail features is considered. The finger-nail biometric is considered as a unique biometric trait using Mel Frequency Cepstral Coefficient (MFCC) technique, finger-knuckle features are extracted and from second level wavelet decomposition the features of finger-nail are extracted. These features are combined using feature level fusion and classified using feedforward backpropagation neural network. Authors in [17] claim feature level fusion require less information to perform the recognition.

3 Result and Discussions

Multimodal fusion techniques: Score level fusion, decision level fusion, and feature level fusion are analyzed considering standard public databases: Biosecure DS-2 [18], FERET [19], VidTIMIT [20], AT&T [21] whose details of the databases are tabulated in Table 1 and standard public databases: USTB I [22], USTB II [22], RUSign, KVKR whose details of the databases are tabulated in Table 2.

Table 1 Details of the databases Biosecure DS-2, FERET, VidTIMIT, AT&T considered for analyses of multimodal biometrics
Full size table
Table 2 Details of the databases USTB I, USTB II, RUSign, KVKR considered for analyses of multimodal biometrics
Full size table

The recognition rate obtained by score level fusion, decision level fusion, and feature level fusion on databases: Biosecure DS-2 [18], FERET [19], VidTIMIT [20], AT&T [21], USTB I [22], USTB II [22], RUsign [23], KVKR [24] are tabulated in Table 3.

Table 3 Recognition rate obtained using score level fusion, decision level fusion, and feature level fusion
Full size table

4 Conclusion

Multimodal systems generally used for face recognition [23–30] can be broadly classified into three categories: Score level fusion, decision level fusion, and feature level fusion. In this paper, we have analyzed the performance of score level fusion, decision level fusion, and feature level fusion on various standard public databases, such as Biosecure DS-2, FERET, VidTIMIT, AT&T, USTB I, USTB II, RUsign and KVKR. From our analysis, we have found that score level fusion approach can effectively fuse multiple biometric modalities, and it is robust to operate in less constrained conditions. Furthermore, score level fusion obtains very accurate performance close to 100 % by restricting the system to accept only high-quality data. In the decision fusion scheme, each decision is made after the improvement of the classifier confidence and hence, the recognition rate obtained is less compared to score level fusion. Feature level fusion requires less information and performs better than decision level fusion, but its recognition rate is less compared to score level fusion. Thus, we conclude that score level fusion is the best fusion technique to recognize images under multimodal biometrics