CNN Implementation for Tourism Accommodation Security via CCTV-Based Facial Recognition

Abstract

Security is an important aspect of tourist accommodation management, particularly for monitoring individuals within the accommodation environment. Closed-Circuit Television (CCTV) can support surveillance activities; however, manual identification remains limited in terms of efficiency and consistency. The application of deep learning, particularly Convolutional Neural Network (CNN) architectures, provides an approach for developing automated face recognition systems. This study aims to analyze and compare the performance of MobileNetV2 and VGG16 architectures for face recognition using images obtained from CCTV recordings in a tourist accommodation environment. The dataset consisted of 1,741 images classified into 10 identity classes, including two staff classes, seven guest classes, and one unknown individual class evaluated under a closed-set scheme. The research process consisted of face detection using Haar Cascade, preprocessing, data augmentation, chronological dataset splitting (sequential temporal split) into training, validation, and testing sets, followed by model training using a transfer learning approach. Model performance was evaluated using accuracy, precision, recall, F1-score, and confusion matrix over repeated experimental runs. The experimental results showed that MobileNetV2 achieved a mean accuracy of 84.00% ± 1.20%, precision of 84.50% ± 1.15%, recall of 84.00% ± 1.20%, and F1-score of 84.10% ± 1.10%. Meanwhile, VGG16 achieved a mean accuracy of 78.00% ± 1.80%, precision of 85.00% ± 1.40%, recall of 78.00% ± 1.80%, and F1-score of 78.10% ± 1.75%. These results indicate that MobileNetV2 provided better identification performance and consistency than VGG16 on the CCTV dataset, particularly in terms of accuracy, recall, and computational efficiency. Therefore, MobileNetV2 was more suitable for the face recognition scenario evaluated in this study.