FE - DI | Dissertações de Mestrado e Teses de Doutoramento
Permanent URI for this collection
Browse
Browsing FE - DI | Dissertações de Mestrado e Teses de Doutoramento by advisor "Alexandre, Luis Filipe Barbosa de Almeida"
Now showing 1 - 5 of 5
Results Per Page
Sort Options
- ForestVISION - Forest Floral Species Classification and ClusteringPublication . Domingos, Gonçalo Gomes; Alexandre, Luis Filipe Barbosa de Almeida; Abreu, António José MarquesThis work aims to solve one of the many tasks responsible for management and monitoring of forests, specifically around Portugal. The task is to classify different species present in the Portugal flora. With the use of artificial intelligence, we were able to successfully distinguish different species present within the images provided. A data set was also created from scratch since there was no public available information regarding the different species around Portugal. Two data sets were created although one of them did not meet the expected results and gave us a lot of problems. However the expected 8 band imagery data set gave us acceptable results. In this thesis, all the steps will be covered with great detail along the chapters, from the build up of the data sets to the real testing of each model tested. We made sure to study the most recent technology and therefore explore the Transformers that have been raising awareness in the Computer Vision field for their characteristics. There is also some ideas for a future development of the respective work, since there can still be some exploration to be made to complete this work.
- LudVision Remote Detection of Exotic Invasive Aquatic Floral Species using Data from a DroneMounted Multispectral SensorPublication . Abreu, António José Marques; Alexandre, Luis Filipe Barbosa de Almeida; Santos, João AmaralRemote sensing is the process of detecting and monitoring the physical characteristics of an area by measuring it’s reflected and emitted radiation at a distance. It is being broadly used to monitor ecosystems, mainly for their preservation. There have been evergrowing reports of invasive species affecting the natural balance of ecosystems. Exotic invasive species have a critical impact when introduced into new ecosystems and may lead to the extinction of native species. In this study, we focus on Ludwigia peploides, considered by the European Union as an aquatic invasive species. Its presence can have negative impacts on the surrounding ecosystem and human activities such as agriculture, fishing, and navigation. Our goal was to develop a method to identify the presence of the species. To achieve this, we used images collected by a dronemounted multispectral sensor. Due to the lack of publicly available data sets containing Ludwigia peploides, we had to create our own data set. We started by carefully studying all the available options. We first experimented with satellite images, but it was impossible to identify the targeted species due to their low resolution. Thus, we decided to use a dronemounted multispectral sensor. Unfortunately, due to budget limitations, we could not acquire the highly specialized types of equipment that is more commonly used in remote sensing. However, we were confident that our setup would be enough to extract the species’ spectral signature, and that the higher resolution compared to satellites would allow us to use deep learning models to identify the species. The use of the drone allowed for better operational flexibility and to cover a large area. The multispectral sensor allowed us to leverage the information of two additional bands outside the visible spectrum. After visiting the study site multiple times and capturing data at various times of the day, we created a representative data set with different atmospheric conditions. After the data collection, we proceeded to the preprocessing and annotation steps to have a usable data set. In later stages, we proved that extracting the specie’s spectral signature from our data set is possible. This was a significant conclusion, as it proved that it is indeed possible to differentiate the species’ spectral signature with equipment that is not as advanced and specialized as the ones used in other studies. After having a data set, we focused on the next step, which was to develop and validate a method that would be able to identify Ludwigia p on our data. We decided on using semantic segmentation models to identify the species. Given that we only have two additional bands compared to traditional RGB images, we could not approach the problem as a standard remote sensing spectroscopy problem. By using semantic segmentation models, we can leverage both the capabilities of these models to recognize objects and the multispectral nature of our data. Fundamentally, the model has the same behavior as usual but has access to the information of two additional bands.We started by using an existing stateoftheart semantic segmentation model adapted to handle our data. After doing some initial tests and establishing a baseline, we proposed and implemented some changes to the existing model. The goal of the modifications was to create a model with lower training times and better performance in detecting Ludwigia p. at high altitudes. The result is a new model better suited to our data and application. Our model is faster when it comes to training time while maintaining similar performance and has a slight performance increase in highaltitude images.
- Machine Learning for the Prediction of App Energy Consumption from Appstore DataPublication . Valente, Daniel Afonso; Alexandre, Luis Filipe Barbosa de AlmeidaThe mobile market has seen tremendous development throughout the past few years both in terms of hardware and the software that is available for the devices. Despite this, the batteries that power these devices have not seen major improvements and have been unable to accompany the progress seen in this field. Due to this phenomenon, researchers have been showing a growing interest in the development of green computing solutions in order to spend the least amount of energy possible when using mobile devices. This as presented itself in a plethora of ways, from the accurate evaluation of the energy consumption of applications through the use of energy models and profilers to the assessment and development of better coding practices with energy conservation as the main focus. However, there have been few to no studies regarding the development of user-side solutions to help solve this problem. In order to fill this gap in research this study focuses on providing a machine learning solution with the intent of identifying links between the information available in the store page of an application and its energy consumption to develop an a priori method for the classification and certification of mobile applications. Hence the main contribution of this project resides on the previously mentioned machine learning model, adapted to the Aptoide appstore and mainly targeting applications that belong to the games category, given that these have the highest volume of downloads and interest by the users of the appstore.
- Redes Neuronais Espaciais e Temporais para a Compreensão de Vídeo em Sistemas EmbebidosPublication . Duarte, Paulo Renato Borges; Alexandre, Luis Filipe Barbosa de Almeida; Neves, João Carlos RaposoA deteção e classificação de ação humana em vídeo são, hoje em dia, tarefas de extrema importância da área de Visão Computacional. Tal importância é atribuída a estas tarefas devido à necessidade de detetar atividade criminosa ou situações de perigo, tornando possível a prevenção e a rápida intervenção no caso de ocorrências das mesmas. Um problema subjacente à utilização desta tecnologia é, precisamente, o elevado poder computacional que lhe está associado, seja a treinar as redes de Aprendizagem Profunda ou na própria inferência. Os dispositivos usados para desempenhar as funções dos sistemas de vigilância são, sobretudo, dispositivos de baixo poder computacional, devido principalmente a fatores como: o elevado custo das placas gráficas e a sua dimensão. É aqui que surgem os problemas que esta dissertação se propõe a tentar resolver. Em virtude da impossibilidade da fase de treino de um modelo ser realizada nos próprios dispositivos e, dado tal processo não ser indispensável, uma vez que esta fase pode ser efetuada em dispositivos com elevado poder computacional, torna-se necessário otimizar o modelo para que este possa ter o menor tempo de inferência e tamanho com a melhor taxa de acertos. Para tentar solucionar este problema, este projeto visa explorar diversas técnicas/métodos de otimização, tais como: fazer uso das camadas convolucionais separáveis, quantização, knowledge distillation, entre outros; assim como criar métodos ou algoritmos que possam ser adicionados ou substituam parte de uma rede.
- A Study on Efficient Semantic SegmentationPublication . Pereira, Luís Carlos Cavaca; Alexandre, Luis Filipe Barbosa de AlmeidaSemantic segmentation extends classical image classification by attributing one class for each pixel in a given image. This approach requires a significant amount of resources to be performed. The majority of time, lowpower resource devices are unable to deliver predictions on this task, because of its computational requirements. Some small robots lack inference speed, enough memory to inference a single instance at time or, even, battery life to delivery continuous predictions. Another aspect, is the incapability of training models on the edge, which can be a major limitation on the practicality of the solution. As if current networks were not big enough for this type of devices, novel architectures tend to be even more complex, which can be seen as a continuous divergence on the possibility of running this kind of models on lowpower devices. With this in mind, the project has the goal of exploring efficient solutions to deploy segmentation models in the edge. To do so, the project aims at exploring efficient architectures and light convolutional layers, alternative segmentation methods and alternative methods of weight representation. In the end, by performing benchmarks on efficient networks with quantization, filter pruning along distillation and layer replacement, it is shown that these methods can be used to save computational resources, but to do so, they sacrifice precision points.