Multi Head Attention with Neural Network Classifier for Face Recognition in Real Time Security Applications

Abstract

Face recognition is an important problem in the fields of pattern recognition, computer vision, AI, and ML. In order to efficiently recognize and differentiate between the photos under consideration, most current face recognition methods focus on identifying the most pertinent facial traits. Most existing face recognition methods take sparse coding at face value, however prototypes of sparse coding are hampered by prohibitively high evaluation and execution costs. Although high-dimensionality is problematic for traditional coding methods, these sparse coding models can handle the complexity of robust face recognition. newlineFirst and foremost, this thesis proposes an ensemble-aided facial recognition strategy that uses an ensemble of feature descriptors and preprocessing methods to achieve strong performance in a natural context. Preprocessed facial photos are mined for a combination of texture and color descriptors, which are then used as inputs to a support vector machine algorithm for classification. Using examples from the FERET and Labeled Faces in the Wild databases, the authors demonstrate the practical results of the proposed methodology. Using additional preprocessing and extracted feature descriptors, the findings reveal that the suggested strategy has good classification accuracy and combination utility of pre-processing techniques. newlineIt is critical to find ways to improve facial recognition that need less computational time and fewer resources. Since that is the case, the second aim of this thesis is to provide new, robust methods for face recognition. ELM classifier is used to recognize high-quality facial photos after they have been preprocessed to obtain multi-scale sparse coding feature patterns. After gathering the sample images used for training the SLFN neural network, the feature patterns dictionaries are used to train the network. The testing results showed that the proposed technique outperformed the SRC, RSC, and MPRSC-ELM classification algorithms in all three scenarios,