Publication: Reconocimiento de objetos en imágenes mediante técnicas de aprendizaje profundo: Redes Neuronales Convolucionales
Loading...
Official URL
Full text at PDC
Publication Date
2019
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
El presente trabajo describe el diseño y desarrollo de una aplicación para el reconocimiento en tiempo real de objetos en imágenes digitales mediante técnicas de aprendizaje profundo (Deep Learning), más concretamente con Redes Neuronales Convolucionales (RNC). Los objetos a reconocer pueden ser de cualquier tipo, si bien en este trabajo se ha orientado hacia el reconocimiento de una plataforma para el aterrizaje seguro de un vehículo aéreo no tripulado VANT (UAV, Unmanned Aerial Vehicle).
La imagen, obtenida por un sistema de captura, se transmite a un servidor capaz de procesar las imágenes utilizando los modelos RNC ya creados y pre-entrenados para responder acorde a la información recibida. Estos modelos se re-entrenan con imágenes propias. Por tanto, el procesamiento de las imágenes consta de las dos fases siguientes:
• Uso de un modelo pre-entrenado, “MobileNet”, especialmente diseñado para su ejecución de manera eficiente en dispositivos integrados, siendo capaz de distinguir entre mil clases de objetos diferentes para realizar una primera predicción.
• A partir del modelo pre-entrenado y con los pesos de las capas de convolución disponibles, se suministran imágenes específicas, objeto del reconocimiento, con el fin de llevar a cabo un proceso de re-entrenamiento, que finaliza con la actualización de los pesos en las mencionadas capas de convolución.
La aplicación consiste en un diseño flexible con un interfaz amigable e intuitiva, que ofrece la posibilidad de ser utilizada sin conocimientos específicos en las tecnologías empleadas. Gracias a esta flexibilidad, es posible utilizar la plataforma para su uso en el reconocimiento de distintos tipos de objetos sin más que realizar el re-entrenamiento con un número relativamente bajo de imágenes que los contienen.
The present work describes the design and development of an application for the recognition in real time of objects in digital images by means of deep learning techniques (Deep Learning), more specifically with Convolutional Neural Networks (CNN). The objects to be recognized can be of any type, but this work has been orientated towards the recognition of a platform for the safe landing of an unmanned aerial vehicle (UAV). The image, obtained by a capture system, is transmitted to a server capable of processing the images using the CNN models already created and pre-trained to respond according to the information received. These models are re-trained with their own images. Therefore, the processing of the images consists of the following two phases: • Use of a pre-trained model, “MobileNet”, especially designed for an efficient execution in integrated devices, being able to distinguish between a thousand different object classes to make a first prediction. • From a pre-trained model and with the weights of the available convolution layers, specific images are provided, object of the recognition, in order to carry out a re-training process, which ends with the updating of the weights in the convolution layers. The application consists of a flexible design with a friendly and intuitive interface, which offers the possibility of being used without specific knowledge in the technologies used. Thanks to this flexibility, it is possible to use the platform for use in recognition of different types of objects without further re-training with a number relatively low of images that contain them.
The present work describes the design and development of an application for the recognition in real time of objects in digital images by means of deep learning techniques (Deep Learning), more specifically with Convolutional Neural Networks (CNN). The objects to be recognized can be of any type, but this work has been orientated towards the recognition of a platform for the safe landing of an unmanned aerial vehicle (UAV). The image, obtained by a capture system, is transmitted to a server capable of processing the images using the CNN models already created and pre-trained to respond according to the information received. These models are re-trained with their own images. Therefore, the processing of the images consists of the following two phases: • Use of a pre-trained model, “MobileNet”, especially designed for an efficient execution in integrated devices, being able to distinguish between a thousand different object classes to make a first prediction. • From a pre-trained model and with the weights of the available convolution layers, specific images are provided, object of the recognition, in order to carry out a re-training process, which ends with the updating of the weights in the convolution layers. The application consists of a flexible design with a friendly and intuitive interface, which offers the possibility of being used without specific knowledge in the technologies used. Thanks to this flexibility, it is possible to use the platform for use in recognition of different types of objects without further re-training with a number relatively low of images that contain them.
Description
Trabajo de Fin de Grado, Universidad Complutense, Facultad de Informática, Departamento de Ingeniería del Software e Inteligencia Artificial, Curso 2018/2019