Multi-ellipse Fitting with PEARL and a Multi-Objective Genetic Algorithm
| Autor: | Heriberto Cruz Hernández, Luis Gerardo de la Fraga |
|---|---|
| Rok vydání: | 2017 |
| Předmět: |
Linear programming
020206 networking & telecommunications 02 engineering and technology Data modeling PEARL (programming language) Data point Local optimum Genetic algorithm Outlier 0202 electrical engineering electronic engineering information engineering 020201 artificial intelligence & image processing Noise (video) Algorithm computer computer.programming_language |
| Zdroj: | ICAPR |
| DOI: | 10.1109/icapr.2017.8592965 |
| Popis: | This paper deals with the multi-ellipse fitting problem, which consist in identifying and fitting an unknown number of ellipses in presence of noise and outliers from data points in R2. This problem is an active research area with many applications in engineering and biology. Numerous studies attempted to solve it by detecting (fitting and extracting) the ellipses in a one by one approach from the data points. Although the one by one approach is effective and useful for many applications, recent studies show that this approach is ill posed and new methods like PEARL have been proposed. PEARL is an energy based multi-model fitting algorithm that considers multiple model instances in the final solution. The PEARL algorithm requires to be initialized with random solutions (models constructed from the minimal number of data points). These solutions are taken by the algorithm and they are refined until obtain a local optimum. Depending on the data points complexity (number of models, level of noise, outliers, and model's spatial support) the number of initial solutions required increases considerably. In this paper we propose the use of a multi-objective genetic algorithm to initialize PEARL. We aim to generate a reduced and high quality initial solutions set to allow PEARL to solve challenging multi-ellipse fitting problem instances. With our approach we are able to solve challenging data points instances, with high amount of noise, up to 100% percentage, and also with overlapping and nested spatial supports. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|