Regresar

Human face orientation based on MobileNetV2 Architecture-Enhanced Convolutional Neural Network

Abstract:

This document details the implementation of software specialized in determining the orientation of the human face in color images. To achieve this, the deep learning convolutional neural network MobileNet-V2 was used as the foundation to develop MobileNetV2 FO. Transfer learning and fine-tuning techniques were employed to estimate the hyper-parameters of MobileNetV2 FO, leveraging prior knowledge to enable the artificial intelligence to specialize in face orientation classification. For the training and inference stages, the PANDORA database was utilized, which contains images of faces in nine orientations. Experimental results demonstrate that MobileNetV2 FO achieves an accuracy of 94.095 %, with precision and sensitivity both at 94.00 %. Furthemore, the AUC (Area Under the Curve) exceeds 94.3% in all cases, and the average precision (AP) is greater than 83 %. This solution provides a robust approach to facial orientation identification, making it an effective, efficient, and versatile tool for various applications in real-world scenarios. It is especially valuable for educational proctoring, biometric security systems, facial ReID, and driver assistance technologies, providing high accuracy and reliability in these crucial areas.