Publication: Algoritmo de detección de intención sexual en imágenes digitales
Loading...
Official URL
Full text at PDC
Publication Date
2021-06
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
Con el rápido avance de la tecnología en los últimos años, el contenido que se difunde en internet y las maneras de hacerlo ha crecido en gran medida. Esta situación se ha hecho accesible para todo tipo de público, haciéndose posible el acceso a información de contenido pornográfico entre niños y jóvenes con mucha facilidad. En el campo del Aprendizaje Automático se vienen haciendo numerosos estudios en cuanto a la detección de la desnudez en las imágenes, con el fin de evitar o minimizar al máximo esta realidad, pues se ha visto que puede causar daños en la estabilidad social y la salud mental de los menores. Sin embargo, en los últimos años se ha comprobado que la detección de la desnudez no es un criterio suficiente para reconocer si una imagen es pornográfica. Entre otros factores, existen también la intención sexual de la imagen. Por este motivo, este trabajo se ha enfocado en la creación de un sistema de detección de intención sexual de la imagen mediante la clasificación de 20 etiquetas para cada una, entre las cuales están: 14 atributos para reconocer desde la postura del cuerpo hasta expresiones faciales, 5 etiquetas para reconocer las emociones y finalmente 1 para saber si la imagen tiene intención sexual, quizás o no. Esta detección se realiza en primer lugar, haciendo uso de técnicas de Visión Artificial para el procesamiento de imágenes. En particular, en cada imagen se extraen las características automáticas siguientes: la capa FC7 de la red neuronal reentrenada CaffeNet y el descriptor de características Histograma Orientado a Gradientes (HOG). Luego, comienza el entrenamiento para el cual se emplean multiplex Máquinas de Soporte Vectorial lineal independientes entre sí para clasificar y además se dispone de un método jerárquico de dos capas. En todo este proceso se ha realizado una investigación exhaustiva, centrando la atención en el Aprendizaje Supervisado y en el Aprendizaje Profundo. Ha sido necesario conocer el trabajo que han realizado otros investigadores, profundizar en este campo, aprender, y también realizar numerosos experimentos. Después de los distintos modelos probados y con los que se ha experimentando haciendo uso de diversos conjuntos de datos que influyen en los resultados del modelo, se ha conseguido un modelo con una exactitud del 63,78 %, que es el que se propone en este Trabajo para detectar la intención sexual.
With the fast progress of technology in recent years, the content that is disseminated on the internet and the ways in which it is disseminated has grown to a large extent. This situation has made it accessible to all kinds of publics, making it possible to access pornographic information for children and young people very easily. In the area of Machine Learning, numerous research projects have been developed on the detection of nudity in images, with the aim of avoiding or minimising this reality as much as possible, as it has been shown that it can cause damage to the social stability and mental health of minors. However, in recent years it has been shown that the detection of nudity is not a sufficient standard for recognising whether an image is pornographic. Besides other factors, there are also the sexual intention of the image. For this reason, this work has focused on the creation of a method to detect the sexual intention of the image by classifying 20 labels for each one, among which are: 14 attributes to recognise from body posture to facial expressions, 5 labels to recognise emotions and finally 1 to know if the image has a sexual intention, possibly or not. This detection is firstly done by making use of Computer Vision techniques for image processing. In particular, the following automatic features are extracted on each image: the FC7 layer of the CaffeNet pre-trained neural network and the Histogram Oriented Gradient (HOG) feature descriptor. Then, training begins for which a number of independent linear Support Vector Machines are used to classify and a two-layer hierarchical method is also available. Throughout this process, exhaustive research has been conducted, with a focus on Supervised Learning and Deep Learning. It has been necessary to learn about the work done by other researchers, to go deeper into the discipline, to learn, and also to experiment extensively. After the different models tested and experimented with using different data sets that influence the results of the model, a model with an accuracy of 63.78 % has been achieved, which is the proposed model in this paper for detecting sexual intent.
With the fast progress of technology in recent years, the content that is disseminated on the internet and the ways in which it is disseminated has grown to a large extent. This situation has made it accessible to all kinds of publics, making it possible to access pornographic information for children and young people very easily. In the area of Machine Learning, numerous research projects have been developed on the detection of nudity in images, with the aim of avoiding or minimising this reality as much as possible, as it has been shown that it can cause damage to the social stability and mental health of minors. However, in recent years it has been shown that the detection of nudity is not a sufficient standard for recognising whether an image is pornographic. Besides other factors, there are also the sexual intention of the image. For this reason, this work has focused on the creation of a method to detect the sexual intention of the image by classifying 20 labels for each one, among which are: 14 attributes to recognise from body posture to facial expressions, 5 labels to recognise emotions and finally 1 to know if the image has a sexual intention, possibly or not. This detection is firstly done by making use of Computer Vision techniques for image processing. In particular, the following automatic features are extracted on each image: the FC7 layer of the CaffeNet pre-trained neural network and the Histogram Oriented Gradient (HOG) feature descriptor. Then, training begins for which a number of independent linear Support Vector Machines are used to classify and a two-layer hierarchical method is also available. Throughout this process, exhaustive research has been conducted, with a focus on Supervised Learning and Deep Learning. It has been necessary to learn about the work done by other researchers, to go deeper into the discipline, to learn, and also to experiment extensively. After the different models tested and experimented with using different data sets that influence the results of the model, a model with an accuracy of 63.78 % has been achieved, which is the proposed model in this paper for detecting sexual intent.
Description
Trabajo de Fin de Grado en Ingeniería Informática, Facultad de Informática UCM, Departamento de Ingeniería del Software e Inteligencia Artificial, Curso 2020/2021