Asdrúbal López‐Chau

Automatic face recognition has evolved significantly with the use of deep learning techniques, enabling more robust systems that can withstand variations in lighting, pose, and expression. This paper presents the implementation of a face recognition system for video recordings using the MTCNN face detection algorithm, which is used for its accuracy in identifying and locating faces in uncontrolled environments with varying lighting conditions. After face detection, the VGG-16 model was used as a deep feature extractor, leveraging its architecture to employ transfer learning techniques to learn hierarchical and discriminative representations of the face. The system processes the video frames using the OpenCV library, then uses the MTCNN network for facial identification, and finally assigns identities through a classifier trained, as previously mentioned, with the VGG-16 model. In recent studies, authors report that convolutional neural networks consistently outperform traditional methods due to their ability to automate the extraction of relevant features and handle complex environmental conditions. Preliminary results indicate that the MTCNN–VGG-16 integration forms an effective strategy for facial recognition in real-world scenarios, demonstrating high accuracy and stability under diverse visual conditions. This work provides an accessible and reproducible architecture for potential applications in video analytics, surveillance, and intelligent automation

La clasificación en conjuntos de datos no balanceados es un problema actual e \nimportante (Khan2012). Ejemplos de aplicaciones del mundo real en las que se \ngeneran este tipo de datos son, por ejemplo, diagnósticos médicos de \nenfermedades tales como cáncer o diabetes mellitus, detección de errores en \ncódigo fuente, ataques informáticos en servidores e identificación de transacciones \nbancarias fraudulentas. En este tipo de escenarios, es sumamente importante \ndetectar casos que generalmente ocurren con poca frecuencia con respecto a la \ncantidad de datos obtenidos, ya que representan un riesgo que hay que observar \noportunamente. Para los ejemplos de aplicaciones mencionadas anteriormente, \nestos casos poco frecuentes serían los pacientes enfermos de cáncer o diabetes, \nlos fragmentos de código fuente con errores, los accesos inválidos a servidores y \nlas transacciones ilícitas. Sin embargo, como se mencionó anteriormente, los \nalgoritmos de clasificación, presentan problemas con conjuntos de datos no \nbalanceados (He2009).