Algoritmo de Estimación de la Edad en Imágenes y Vídeos
Loading...
Official URL
Full text at PDC
Publication date
2021
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citation
Abstract
Las técnicas para la detección de la edad pueden ser determinantes para personas que no se encuentren identificadas y se necesite saber la edad por diversos motivos, como delitos, abusos sexuales, etc, ya que la intervención humana en estos casos se trata de una tarea muy lenta y ralentizarían todo el proceso. Es por esto que perfeccionar las actuales técnicas de estimación de la edad, puede constituir una mejora para los ámbitos forenses y de seguridad. Este trabajo puede resultar un progreso en este ámbito, ya que muchos delitos son cometidos con nocturnidad y procesar las imágenes con filtros antes de introducirlas en un modelo de detección, podrían ser resultantes. Por ello, en este trabajo se muestran los objetivos obtenidos de aplicar preprocesamiento a las imágenes antes de introducirlas en un modelo de detección. Después de una ardua investigación en la que ha habido momentos complicados, como la búsqueda de un buen equipo para poder
realizar la ejecución del modelo y el procesamiento de todas las imágenes del dataset, manejar de manera fluida TensorFlow, etc, se consigue solucionar mediante la creación
de una instancia en Google Cloud. Una vez conseguido ésto, se procede a la realización de un estudio sobre la mejora de la calidad de las imágenes relacionadas con los datasets
mencionados en este trabajo. Con este análisis, se pretende abordar los problemas que pueden ocasionar la mala calidad de las imágenes, así como, su mal tratamiento a la hora de introducirlas en un modelo. Para el desarrollo de este estudio se ha investigado sobre distintos filtros para la mejora de imágenes, ejecutándose y viendo el desempeño de
cada uno y guardando los mejores resultados, para su posterior elección. También, se han investigado otras formas de detección del rostro, como la detección de perfil. Tras reunir todo el preprocesamiento con mejores resultados e introducirlo en un modelo para ver las mejoras respecto a las imágenes sin tratar, sale como vencedor el modelo entrenado con aquellas imágenes preprocesadas anteriormente, obteniéndose una mejora en la detección de la edad.
Age detection techniques can be decisive for people who are not identified and age needs to be known for various reasons, such as crimes, sexual abuse, etc., since human intervention in these cases is a very difficult task. slow and slow down the whole process. This is why perfecting current age estimation techniques can be an improvement for forensic and security fields. This work can be a progress in this area, since many crimes are committed at night and processing the images with filters before introducing them into a detection model could be the result. For this reason, this work shows the objectives obtained from applying preprocessing to the images before introducing them into a detection model. After an arduous investigation in which there have been complicated moments, such as the search for a good team to be able to carry out the execution of the model and the processing of all the images of the dataset, to handle TensorFlow in a fluid way, etc., it is possible to solve by means of the creating an instance in Google Cloud. Once this has been achieved, a study is carried out on improving the quality of the images related to the datasets mentioned in this work. With this analysis, it is intended to address the problems that can cause the poor quality of the images, as well as their poor treatment when introducing them into a model. For the development of this study, different filters have been investigated for the improvement of images, executing and seeing the performance of each one and saving the best results, for later selection. Also, other forms of face detection have been investigated, such as profile detection. After gathering all the preprocessing with better results and introducing it into a model to see the improvements compared to the untreated images, the model trained with those images previously preprocessed was the winner, obtaining an MAE of 0.24 %.
Age detection techniques can be decisive for people who are not identified and age needs to be known for various reasons, such as crimes, sexual abuse, etc., since human intervention in these cases is a very difficult task. slow and slow down the whole process. This is why perfecting current age estimation techniques can be an improvement for forensic and security fields. This work can be a progress in this area, since many crimes are committed at night and processing the images with filters before introducing them into a detection model could be the result. For this reason, this work shows the objectives obtained from applying preprocessing to the images before introducing them into a detection model. After an arduous investigation in which there have been complicated moments, such as the search for a good team to be able to carry out the execution of the model and the processing of all the images of the dataset, to handle TensorFlow in a fluid way, etc., it is possible to solve by means of the creating an instance in Google Cloud. Once this has been achieved, a study is carried out on improving the quality of the images related to the datasets mentioned in this work. With this analysis, it is intended to address the problems that can cause the poor quality of the images, as well as their poor treatment when introducing them into a model. For the development of this study, different filters have been investigated for the improvement of images, executing and seeing the performance of each one and saving the best results, for later selection. Also, other forms of face detection have been investigated, such as profile detection. After gathering all the preprocessing with better results and introducing it into a model to see the improvements compared to the untreated images, the model trained with those images previously preprocessed was the winner, obtaining an MAE of 0.24 %.
Description
Trabajo de Fin de Grado en Ingenieria del Software, Facultad de Informática UCM, Departamento de Ingeniería de Software e Inteligencia Artificial, Curso 2020/2021.