Desarrollo e implementación de un sistema de control de asistencia de estudiantes basado en reconocimiento facial”
No Thumbnail Available
Date
2024-10
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad Nacional de Trujillo
Abstract
Este proyecto de ingeniería tiene como objetivo principal desarrollar e implementar un
sistema de control de asistencias de estudiantes basado en reconocimiento facial en la escuela de
Ingeniería Mecatrónica. En primer lugar, se procedió a determinar los requerimientos del sistema,
el cual, en resumen, consta de un funcionamiento rápido, eficiente y robusto. Luego se procedió a
obtener base de datos de rostros, para lo cual en este caso se usaron las datasets Labeled Faces in
the Wild(LFW) y FEI Database, debido a que la primera es la más habitual al momento de entrenar
redes siamesas convolucionales y la segunda debido a la similitud que tiene con el ambiente en el
cual el sistema fue implementado. Posteriormente se procedió a la selección de una arquitectura
de reconocimiento facial, para lo cual se tomaron en cuenta la precisión, la robustez, la capacidad
de usuario, la facilidad de registro y el número de imágenes necesarias para el reconocimiento. Se
entrenaron en total 4 modelos pre entrenados (ResNet50, Xception, SeReSNext50 y Inception
ResNet V2), dos arquitecturas (Red Siamesa Convolucional y Red Siamesa Convolucional con
Pérdida Triple) y dos datasets (LFW, FEI), dando un total de 16 entrenamientos. El modelo con
mejor resultado fue la red siamesa convolucional con pérdida triple basado en ResNet50 usando
la dataset LFW, presentando una exactitud de modelo de 71.87%. Luego se realizó la
implementación de una aplicación integrada con interfaz gráfica (GUI), que, usando el modelo
obtenido, realiza la toma de asistencia, así como el registro de alumnos nuevos y la exportación de
datos en formato Excel. Finalmente, este sistema obtuvo una precisión de un 92.5% y un tiempo
de respuesta promedio de 3.3 segundos.
The main objective of this engineering project is to develop and implement a student attendance control system based on facial recognition in the school of Mechatronics Engineering. First, we proceeded to determine the requirements of the system, which, in summary, consists of a fast, efficient and robust operation. Then we proceeded to obtain the face database, for which in this case the Labeled Faces in the Wild (LFW) and FEI Database datasets were used, since the former is the most common when training convolutional Siamese networks and the latter due to its similarity with the environment in which the system was implemented. Subsequently, a face recognition architecture was selected, taking into account accuracy, robustness, user capacity, ease of registration and the number of images required for recognition. A total of 4 pre-trained models (ResNet50, Xception, SeReSNext50 and Inception ResNet V2), two architectures (Convolutional Siamese Network and Convolutional Siamese Network with Triple Loss) and two datasets (LFW, FEI) were trained, giving a total of 16 trainings. The best performing model was the convolutional Siamese network with triple loss based on ResNet50 using the LFW dataset, presenting a model accuracy of 71.87%. Then, an integrated application with a graphical user interface (GUI) was implemented, which, using the model obtained, performs the attendance taking, as well as the registration of new students and the export of data in Excel format. Finally, this system obtained an accuracy of 92.5% and an average response time of 3.3 seconds.
The main objective of this engineering project is to develop and implement a student attendance control system based on facial recognition in the school of Mechatronics Engineering. First, we proceeded to determine the requirements of the system, which, in summary, consists of a fast, efficient and robust operation. Then we proceeded to obtain the face database, for which in this case the Labeled Faces in the Wild (LFW) and FEI Database datasets were used, since the former is the most common when training convolutional Siamese networks and the latter due to its similarity with the environment in which the system was implemented. Subsequently, a face recognition architecture was selected, taking into account accuracy, robustness, user capacity, ease of registration and the number of images required for recognition. A total of 4 pre-trained models (ResNet50, Xception, SeReSNext50 and Inception ResNet V2), two architectures (Convolutional Siamese Network and Convolutional Siamese Network with Triple Loss) and two datasets (LFW, FEI) were trained, giving a total of 16 trainings. The best performing model was the convolutional Siamese network with triple loss based on ResNet50 using the LFW dataset, presenting a model accuracy of 71.87%. Then, an integrated application with a graphical user interface (GUI) was implemented, which, using the model obtained, performs the attendance taking, as well as the registration of new students and the export of data in Excel format. Finally, this system obtained an accuracy of 92.5% and an average response time of 3.3 seconds.
Description
Keywords
TECHNOLOGY::Engineering mechanics