FaceNet-MMAR: Revolutionizing Facial Recognition for Smart University Libraries

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Jan 18 2024

In a paper published in the journal PLOS ONE, researchers detailed integrating facial recognition technology (FR-tech) into constructing innovative libraries in higher education. The study optimized traditional FaceNet algorithms, incorporating enhancements like mobilenet, attention mechanism, receptive field module, and mish activation function.

The resulting multitask face recognition convolutional neural network demonstrated outstanding performance. The model received high satisfaction from teachers and students, showcasing its efficacy in providing robust technical support for developing intelligent university libraries.

Related Work

Past research has highlighted computer vision, particularly FR-tech, drawing attention to its versatile applications. Recent studies have delved into constructing intelligent facial recognition models in university libraries, addressing challenges posed by varying lighting and facial conditions. Innovations include methods proposed by researchers, real-time systems, and novel algorithms. Researchers have actively explored advancements in adversarial networks, convolutional neural network (CNN) structures, and age-aware models. Despite these efforts, the application of facial recognition in intelligent university libraries still needs to be explored.

Smart Library FR

The paper addresses challenges traditional FaceNet networks face, including large parameter quantities, complex calculations, and high model memory consumption, hindering accurate FR in university intelligent libraries. Researchers initiate the study by scrutinizing the functional modules of the library access control system, paving the way for the design and implementation of an intelligent management system grounded in the uncovered insights.

The proposed system comprises six modules: system management, personnel management, equipment management, FR, access management, and external empowerment. It addresses issues with data, space, resource management, user experience, and traffic management.

The hardware construction of the smart library system involves equipment acquisition, communication networks, and server layers. The study structures the network topology to ensure the normal and orderly operation of the system. The improved FaceNet face recognition algorithm integrates the multitask CNN (MTCNN) for facial detection, with three sub-networks (proposal network (P-Net), refine network (R-Net), and output network (O-Net) enhancing candidate windows and feature point localization. Following the MTCNN algorithm, the FaceNet network actively engages in FR. However, traditional FaceNet networks face limitations, leading the study to propose the FaceNet-MMAR model, optimizing with mobilenet, mish activation function, attention module, and receptive field module.

The attention module enhances the feature extraction ability of the backbone network, and the receptive field block (RFB) module in the final layer of the leading feature network improves the image receptive field. The construction of the FaceNet-MMAR model actively incorporates the mish activation function and introduces enhancements to the attention and RFB modules.

The attention module utilizes weighted input image features to improve feature extraction, while the RFB module expands the receptive field through various expansion rates and convolution kernel sizes. Applying the FaceNet-MMAR model to the access control system of university libraries is anticipated to enhance FR accuracy and efficiency in real-world scenarios.

FaceNet-MMAR: Superior Recognition Performance

The study begins by comparing the performance of various FR algorithms, focusing on feature matching error, loss curve variation, receiver operating characteristic (ROC) curve, and recognition performance. The FaceNet-MMAR algorithm performs excellently, showcasing its superiority in feature matching error, loss curve stability, ROC curve analysis, and overall recognition accuracy.

The experimental setup involves essential hardware components, such as a high-quality camera and a powerful computer. It employs the TensorFlow framework, python programming language, open-source computer vision (OpenCV) image processor, and training datasets. The FaceNet-MMAR algorithm is tested against the Chinese Academy of Sciences Institute of Automation webface (CASIA-WebFace) dataset, demonstrating superior feature matching, loss curve stability, and ROC curve results compared to the FaceNet-mobilenet.

The FaceNet-MMAR algorithm performs a comprehensive evaluation to assess its recognition accuracy, demonstrating its performance across training and validation datasets. Feature matching errors, loss curves, and ROC curves are analyzed, revealing that FaceNet-MMAR outperforms FaceNet-mobilenet regarding accuracy, stability, and area under the ROC curve. The study compares recognition performance across different models, emphasizing the FaceNet-MMAR model's superiority with a recognition accuracy of 99.05% and a minimal recognition error of 0.51%.

The study actively assesses the application of the FaceNet-MMAR model in the FR system of the university's innovative library. The study splits facial data from a university's innovative library into training and validation sets. Different FR methods' recognition accuracy and recall rates, including eigenface, local binary pattern (LBP), fisherface, and FaceNet-MMAR, are compared. FaceNet-MMAR consistently outperforms the other methods, exhibiting higher accuracy, recall rates, and stable recognition times. Moreover, satisfaction statistics from university teachers and students highlight FaceNet-MMAR's superior performance, achieving 97.6% teacher satisfaction and 96.8% student satisfaction.

Finally, the visualization analysis results demonstrate FaceNet-MMAR's effective FR in an intelligent library setting. The model identifies distinct facial features, ensuring accurate and efficient access control recognition. The study concludes that the FaceNet-MMAR algorithm provides exceptional facial recognition performance, making it well-suited for practical applications in smart-university libraries.

Conclusion

In summary, this study enhances the FaceNet network with attention mechanisms and RFB to improve FR accuracy and efficiency. FaceNet-MMAR outperforms FaceNet-MN, achieving a minimal error of 0.04 and maintaining a stable loss curve during iterations. The model exhibits superior FR accuracy (99.05%) compared to other traditional algorithms, showcasing its practical application in innovative university library systems. FaceNet-MMAR demonstrates efficient recognition times, earning high satisfaction rates (97.6% for teachers and 96.8% for students). While victorious, testing errors may arise from limited data samples, suggesting the potential for further exploration with more extensive datasets.