A Robust Explainable Quantum Ensemble Deep Learning Architecture with Optimized Feature Selection For 3d Face Recognition

Abstract

In computer vision, face recognition is regarded as one of the most significant applications. and AI owing to its broad usage in security systems, surveillance systems, biometric recognition systems, and even smart human-computer interactions. However, factors such as changes in illumination, facial expression, pose, occlusion, and image quality still create huge barriers in performing accurate face recognition. In order to overcome the above issues, this paper proposes an Explainable Stacking Ensemble Quantum Deep Learning (ESEQDL) along with Intelligent Feature Selection scheme for face recognition using graph neural network (GNN) landmarks that create three-dimensional (3D) face meshes from two-dimensional (2D) facial images from different sources. Preprocessing makes use of Restormer to improve image quality, followed by segmentation through the deep learning approach using the SegFormer framework that helps in separating the foreground from the background and thus making the detection more accurate and noise-free. Feature Extraction in the Vision Transformer (ViT) architecture. Additionally, a feature selection method is implemented to help select the most important facial features and exclude unnecessary information. For the purpose of this study, Optimized Recursive Feature Elimination (ORFE) technique is used for feature selection. The above proposed framework uses several deep learning algorithms into one stacking ensemble quantum algorithm for better feature extraction and performance of classification tasks. In this study, the different types of explainable ensemble quantum deep neural networks used include SwinFace, S-ViT, InceptionNet QCNN, Grad CAM and SHAP for feature extraction of diverse and discriminating facial representations from the input face image. The meta-classifier uses the output of all different basis models. This approach increases the resistance and generalization capacity of the face recognition system through the use of the strengths of multiple architectures of deep learning. Evaluation of the proposed approach is done on several face images datasets in various environmental conditions. The results of the experiment readily demonstrate that the stacking standard deep quantum learning approach coupled with intelligent feature selection performs better than the traditional systems.