Titulo Estágio
Neuroevolution of efficient models for image segmentation
Áreas de especialidade
Sistemas Inteligentes
Local do Estágio
DEI-FCTUC
Enquadramento
Humans use vision as their primary sense to perceive the surrounding environments. Over the years, researchers have been proposing various computer vision methods to attempt to provide the same capabilities to robots. A popular computer vision problem is image segmentation, which attempts to identify an object in an image as well as its location. Image segmentation has various applications, particularly in the medical domain (e.g., identifying necrotic tissue) or autonomous navigation (e.g., selecting a clear path to move on). Currently, the best performance is achieved by approaches based on Deep Neural Networks that require extensive computational resources. Such computational requirements are the main drawback of these approaches, making them inadequate to be deployed in robotics and handheld devices, where there are limited resources and concerns regarding power efficiency. Moreover, if such models are to be used for real time operation, the need for high efficiency is further increased.
Objetivo
The main goal of this dissertation is to improve the existing neuroevolution algorithms to automatically design memory and energy efficient models for performing image segmentation in real time in resource constrained systems. Neuroevolution consists of the application of Evolutionary Algorithms to automatically design Artificial Neural Networks (ANN). This is a highly relevant problem, as the architecture of ANNs greatly influences their performance and their design is typically a manual trial-and-error process. However, most literature focuses on evolving ANNs to maximize performance, with no regard to their efficiency. This dissertation aims to optimize ANNs for both performance and efficiency, building on the Bio-Inspired Artificial Intelligence group (bAI) significant experience in the field of neuroevolution (e.g., DENSER [1]). The methods devised shall be tested against those of the state of the art in popular benchmark problems.
[1] -Assunção, F., Lourenço, N., Machado, P., & Ribeiro, B. (2019). DENSER: deep evolutionary network structured representation. Genetic Programming and Evolvable Machines, 20, 5-35.
Plano de Trabalhos - Semestre 1
- Literature review.
- Implementation and test of the most promising approaches.
- Development of a neuroevolution approach to maximize performance in image segmentation.
- Writing of the intermediate report.
Plano de Trabalhos - Semestre 2
- Development of neuroevolution approaches for simultaneous optimisation of performance and efficiency in image segmentation.
- Validation of the proposed methods and comparison with the existing approaches.
- Writing of the dissertation.
- Writing of a scientific article with the main results.
Condições
The work shall be carried out within CISUC’s Bio-Inspired Artificial Intelligence group (bAI), under the supervision of Prof. João Macedo and Prof. Nuno Lourenço. Additionally, eligible students may have the opportunity to receive a scholarship (Bolsa de Investigação para Licenciado) following the monthly stipend guidelines set by Fundação para a Ciência e Tecnologia (FCT).
Orientador
João Macedo / Nuno Lourenço
jmacedo@dei.uc.pt 📩