Extreme gradient boosting and fractional attention network for three-dimensional biometric face reconstruction and recognition

The Three-Dimensional biometric face recognition aims to recreate human faces using Two-Dimensional images and identify them precisely based on their facial features. This technology is essential for robust identity verification, particularly in various expression conditions, poses, and changes in lighting. Different traditional approaches utilized to recognize the face struggled to detect the small variations in facial features, which led to missed identifications and high false-positive rates. Hence, a new model called eXtreme Gradient Boosting-Fractional Attention Network is introduced for Three-Dimensional biometric face reconstruction and recognition from Two-Dimensional images. Initially, the input image is pre-processed by exploiting the Adaptive Median Filter. After this, the Three-Dimensional face is reconstructed from the Two-Dimensional face image using the Dual Attention MobDenseNet with a modified loss function. Next, features including Linear Discriminant Analysis of surface normals, depth images, shape index, surface normal, and point cloud representation are mined. Lastly, the face is recognized by exploiting the eXtreme Gradient Boosting-Fractional Attention Network technique. Further, the eXtreme Gradient Boosting-Fractional Attention Network is formulated by combining Convolutional eXtreme Gradient Boosting, Neuron Attention Stage by Stage Network, and Fractional Calculus. The evaluation of the devised model is done by using the Surveillance Cameras Face Database and Masked Face Recognition Dataset based on the hold-out method and cross-validation. Compared to existing methods, the newly presented eXtreme Gradient Boosting-Fractional Attention Network approach measured a maximum specificity of 95.999%, sensitivity of 97.888%, and accuracy of 96.998%.