Publication: Detección facial en vídeos digitales
Loading...
Official URL
Full text at PDC
Publication Date
2020-06
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
Hoy en día, las redes neuronales constituyen un pilar fundamental en el desarrollo de aplicaciones inteligentes. No obstante, dicho campo se encuentra aún en fase de expansión y son muchos los avances que se consiguen diariamente. La inteligencia artificial es la responsable de crear sistemas autónomos y de automatizar tareas que suponen una gran inversión de tiempo y esfuerzo de ser realizadas por seres humanos, como la identificación de objetos tanto en imágenes como en vídeos. Por otro lado, el análisis forense digital ha sufrido un avance considerable en los últimos años gracias a la evolución tecnológica continua que se está viviendo actualmente. Si se juntan la inteligencia artificial y el análisis forense, se puede mejorar considerablemente dicha labor, y permitir así abrir las puertas a futuros avances que den paso a funcionalidades que aún están por conocer.
Este trabajo se centra en intentar crear sistemas óptimos especializados en encontrar rostros en vídeos con condiciones muy desfavorables, con el fin de encontrar comportamientos delictivos de manera rápida y eficiente, facilitando así la labor del análisis forense y desprenderse así de la necesidad de analizar los vídeos manualmente.
Para ello, se ha realizado una extensa investigación tanto en los conceptos que engloban el campo de la detección facial, como en trabajos previos relacionados con este tema, para comprender en profundidad cuáles son las técnicas actuales en uso que podrían servir para este trabajo.
Tras realizar una amplia experimentación probando con seis modelos distintos basados en aprendizaje profundo, se ha llegado a la conclusión de proponer dos modelos: uno especializado en la detección de vídeos en local capaz de inferir con una exactitud del 74,83 %, y otro adaptado para su funcionamiento en tiempo real con el que se alcanza una exactitud del 74,59%.
Today, neural networks are a fundamental pillar in the development of intelligent applications. However, this field is still in a phase of expansion and many advances are being made every day. Artificial intelligence is responsible for creating autonomous systems and for automating tasks that require a great investment of time and effort to be carried out by human beings, such as the identification of objects in both images and videos. On the other hand, digital forensic analysis has undergone a considerable advance in recent years thanks to the continuous technological evolution that is currently taking place. If artificial intelligence and forensic analysis are brought together, this work can be considerably improved, thus opening the door to future advances that will give way to functionalities that are still unknown. This work is focused on trying to create optimal systems specialized in finding faces in videos with very unfavorable conditions, in order to find criminal behavior quickly and efficiently, thus facilitating the work of forensic analysis and thus get rid of the need to analyze the videos manually. For this purpose, extensive research has been carried out both in the concepts that encompass the field of facial detection, as well as in previous work related to this topic, in order to understand in depth which are the current techniques in use that could serve for this work. After carrying out extensive experimentation by testing six different models based on deep learning, the conclusion has been reached that two models should be proposed: one specialized in the detection of videos in premises capable of inferring with an accuracy of 74.83 %, and another adapted for real time operation with which an accuracy of 74.59% is achieved.
Today, neural networks are a fundamental pillar in the development of intelligent applications. However, this field is still in a phase of expansion and many advances are being made every day. Artificial intelligence is responsible for creating autonomous systems and for automating tasks that require a great investment of time and effort to be carried out by human beings, such as the identification of objects in both images and videos. On the other hand, digital forensic analysis has undergone a considerable advance in recent years thanks to the continuous technological evolution that is currently taking place. If artificial intelligence and forensic analysis are brought together, this work can be considerably improved, thus opening the door to future advances that will give way to functionalities that are still unknown. This work is focused on trying to create optimal systems specialized in finding faces in videos with very unfavorable conditions, in order to find criminal behavior quickly and efficiently, thus facilitating the work of forensic analysis and thus get rid of the need to analyze the videos manually. For this purpose, extensive research has been carried out both in the concepts that encompass the field of facial detection, as well as in previous work related to this topic, in order to understand in depth which are the current techniques in use that could serve for this work. After carrying out extensive experimentation by testing six different models based on deep learning, the conclusion has been reached that two models should be proposed: one specialized in the detection of videos in premises capable of inferring with an accuracy of 74.83 %, and another adapted for real time operation with which an accuracy of 74.59% is achieved.
Description
Trabajo de Fin de Grado en Ingeniería Informática y Grado en Ingeniería del Software, Facultad de Informática UCM, Departamento de Ingeniería del Software e Inteligencia Artificial, Curso 2019/2020.