International Journal of Science, Engineering and TechnologyAuthors: Kanupriya Upadhyay, Manisha Sharma
Abstract: In settings where resources are limited, traditional automated attendance systems have a lot of trouble because they are slow to respond and are sensitive to changes in the environment. Li et al. research presents an enhanced facial recognition framework that mitigates the efficiency deficit in lightweight segmentation as identified [1] . Ryando et al. pointed out To improve security against the spoofing vulnerabilities [2] in healthcare kiosks, we add a liveness detection module that is based on the Eye Aspect Ratio (EAR) looked at heavy deep learning architectures that put a lot of strain on computers. In contrast, our system uses an Adaptive Circular Kernel Extreme Learning Machine (ACK-ELM) to achieve a non-iterative, one- shot learning paradigm. This method is based on the fast hybrid ideas of Anil and Suresh [4] and combines Histogram of Oriented Gradients (HOG) with shallow MobileNetV2 features. To ensure resilience against the lighting and expression varia- tions examined by Abdallah et al., the model is validated on the Extended Yale B dataset. The practical deployment phase comes after the five-phase attendance marking architecture that Potdar et al. [6] proposed. It uses a MySQL database to keep track of records in real time. Our classification strategy also works better than the Support Vector Machine (SVM) baselines set by Ali et al.[8] because it makes inferences faster. By using the few-shot learning efficiencies that Nasralla [9] looked into in the AIFS framework, the system stays very accurate even when it doesn’t have a lot of training data. To make the best use of memory, we use the Global Average Pooling (GAP) techniques suggested by Wei et al. [9]. These techniques compress features to keep the system from crashing. Lastly, the system uses the multimodal fusion logic of Abdul-Al et al. [10] and the temporal consistency principles of Interno` et al. [11] to tell the difference between real facial trajectories and fake or static ones. Experimental results show that the accuracy is 97.16% on a standard CPU, which makes it a useful solution for large-scale institutional attendance.
DOI: http://doi.org/10.5281/zenodo.20156258
