Biometric identification of humans has many compelling advantages, but some challenges must be overcome. Each biometric modality (face, fingerprint, iris, and so on) has some number of individuals for which it is difficult to apply or inaccurate in performance. For example, some occupations or health conditions can make high-quality fingerprints difficult to acquire reliably. One strategy for dealing with this is the use of multimodal biometric systems, in which more than one type of biometric sample is acquired and used for identification. The benefit is obvious, but the new challenge is to find proper methods for combining the several modalities. Proposals range from decision-level fusion, where the final results of each individual identification attempt are examined, back to feature-level fusion, where a combined mathematical representation of multiple biometric modalities is used for classification.

In this paper, Saini and Sinha compare feature-level and score-level fusion for a specific method of extracting features from biometric samples in two modalities: face and palmprint. They describe the problem in detail, give a summary of existing work in multimodal biometric fusion, and then present their feature extraction technique, based on the Gabor-Wigner transform of image data. Choice of feature extraction technique is important since feature-level fusion is being tested; clearly, the feature extraction must be well suited for all modalities being used. As they discuss, the Gabor-Wigner transform is a combination of two mathematical filters used for the simultaneous representation of location and spatial frequency information. The Gabor filter can be thought of as the product of a Gaussian envelope and a directed 2D cosinusoid; the resulting response at a given position, frequency, and angle indicates local spatial frequency activity. The Wigner transform also represents frequency content at a location, in x and y components. Both transforms have drawbacks: the Gabor suffers from low resolution, while the Wigner has a nonlinear cross term. The combined Gabor-Wigner approach is an attempt to mitigate the limitations of the two separate filters.

Saini and Sinha present a feature extraction system in which the 2D Gabor transform of an image is calculated, followed by application of a 1D Wigner filter. The result is an image with a number of filter response values at each location. To deal with the large number of results, the authors limit their image size to 50 by 50 locations, and keep 25 response values at each location, to produce a feature vector of size 62,500. Their choice of using a 1D Wigner filter rather than a 2D one seems limiting; some justification for this decision would add greatly to the paper.

Next, these feature vectors are used in two types of decision engine. In one case, the two feature vectors (one for palmprint and one for face) are combined using a weighted sum to produce a complex (2D) result: the face features are the real component and the palmprint features are used as the imaginary component. On these complex vectors, classification is performed using particle swarm optimization. The weight values are selected by an optimization process to produce the highest accuracy. Additionally, each feature vector is submitted to a separate particle swarm optimization classifier. The separate classification results (matching scores) are weighted and fused to yield a combined score. To evaluate the advantage in each approach, a hybrid system is devised where all three scores--the result of the separate classifiers as well as the combined classifier--are fused using a sum rule, with weights applied to each score. The authors examine the resulting scores to determine that the score-level fusion outperforms the feature-level fusion, but that the hybrid system performs best of all. Notably, the optimal weights found for the three scores are generally close to one-third, indicating that no one approach is dominating the score fusion.

This paper describes a system using a hybrid method of fusing multiple biometric modalities, and shows improved accuracy from the approach. Some of the findings can be expected to be general (feature fusion requires suitable feature extraction methods, score fusion of individual and hybrid classifiers is a good strategy), while others may be specific to the images and details of the method. In addition, the technique could stand testing over many more images. Nonetheless, this is an interesting implementation of multimodal biometric fusion.