Multimodal biometrics, using machine and deep learning, has recently gained interest over single biometric modalities. This interest stems from the fact that this technique improves recognition and, thus, provides more security. In fact, by combining the abilities of single biometrics, the fusion of two or more biometric modalities creates a robust recognition system that is resistant to the flaws of individual modalities. However, the excellent recognition of multimodal systems depends on multiple factors, such as the fusion scheme, fusion technique, feature extraction techniques, and classification method.
In machine learning, existing works generally use different algorithms for feature extraction of modalities, which makes the system more complex. On the other hand, deep learning, with its ability to extract features automatically, has made recognition more efficient and accurate. Studies deploying deep learning algorithms in multimodal biometric systems tried to find a good compromise between the false acceptance and the false rejection rates (FAR and FRR) to choose the threshold in the matching step. This manual choice is not optimal and depends on the expertise of the solution designer, hence the need to automatize this step. From this perspective, the second part of this thesis details an end-to-end CNN algorithm with an automatic matching mechanism.
This thesis has conducted two studies on face and iris multimodal biometric recognition. The first study proposes a new feature extraction technique for biometric systems based on machine learning. The iris and facial features extraction is performed using the Discrete Wavelet Transform (DWT) combined
with the Singular Value Decomposition (SVD). Merging the relevant characteristics of the two modalities is used to create a pattern for an individual in the dataset. The experimental results show the robustness of our proposed technique and the efficiency when using the same feature extraction technique for both modalities. The proposed method outperformed the state-of-the-art and gave an accuracy of 98.90%.
The second study proposes a deep learning approach using DensNet121 and FaceNet for iris and faces multimodal recognition using feature-level fusion and a new automatic matching technique. The proposed automatic matching approach does not use the threshold to ensure a better compromise between performance and FAR and FRR errors. However, it uses a trained multilayer perceptron (MLP) model that allows people’s automatic classification into two classes: recognized and unrecognized. This platform ensures an accurate and fully automatic process of multimodal recognition. The results obtained by the DenseNet121-FaceNet model by adopting feature-level fusion and automatic matching are very satisfactory. The proposed deep learning models give 99.78% of accuracy, and 99.56% of precision, with 0.22% of FRR and without FAR errors.
The proposed and developed platform solutions in this thesis were tested and vali- dated in two different case studies, the central pharmacy of Al-Asria Eye Clinic in Dubai and the Abu Dhabi Police General Headquarters (Police GHQ). The solution allows fast identification of the persons authorized to access the different rooms. It thus protects the pharmacy against any medication abuse and the red zone in the military zone against the unauthorized use of weapons.