Segmentação interativa de imagens usando redes complexas e competição e cooperação entre partículas
Carregando...
Data
2019-01-30
Autores
Passerini, Jefferson Antonio Ribeiro
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
A segmentação de imagens é o processo de identificar e separar estruturas e objetos relevantes em uma imagem, não é uma tarefa trivial para um algoritmo computacional devido à complexidade dos elementos envolvidos no processo. A variedade de imagens e sua grande quantidade de características impossibilitam a utilização de apenas um modelo para a tarefa de segmentação. Deste modo, a aplicação de algoritmos de aprendizado de máquina, ganha importância, pois este ramo da inteligência artificial tenta reproduzir a forma de “aprender” humano. A aplicação do modelo de competição e cooperação entre partículas na segmentação de imagens mostra-se uma abordagem interessante pois possui baixo custo computacional, além de reduzir a necessidade de especialistas para a geração de dados rotulados. Este trabalho propõe melhorias na construção da rede complexa que o modelo de competição e cooperação entre partículas utiliza-se, através da alteração das características extraídas dos píxeis, mudança na forma de ligação dos vértices do modelo utilizando a similaridade das características e a vizinhança espacial dos píxeis da imagem e a influência gerada por uma partícula sob outros vértices vizinhos dentro da rede complexa. A metodologia proposta resultou em uma melhoria do nível de automação do processo, eliminando a necessidade de incrementar ou decrementar a importância de uma característica sob as outras. Comparativamente, durante os testes realizados com 151 imagens e 4530 amostras de execução para cada imagem em cada modelo, a abordagem proposta obteve uma taxa de erro de 0,49% na classificação dos píxeis contra 3,14% para o método testemunha, sem a utilização de ajustes, além de mostrar-se mais estável em relação a este desempenho perante as diferentes imagens testadas, e ainda obteve melhora no tempo final de processamento do modelo.
Image segmentation is the process of identifying and separating relevant structures and objects in an image, it is not a trivial task for a computational algorithm due to the complexity of the elements involved in the process. The variety of images and its vast number of characteristics make it impossible to use only a model for the segmentation task. In this way, the application of machine learning algorithms is highlighted, because this branch of artificial intelligence tries to reproduce the way of human learning. The application of the competition and cooperation model among particles in the image segmentations is considered an interesting approach since it has low computational cost, besides reducing the need of specialists for generating labeled data. This work proposes improvements in the making of the complex network that the competition and cooperation model among particles uses, through the alteration of the characteristics extracted from pixels, changes in the shape of the model’s connection of the vertices using the similarity of the characteristics and the image pixel’s spatial neighboring and the influence generated by a particle under the other neighbor vertices within the complex network. The proposed methodology resulted in an improvement of the process’s automation level, eliminating the need to increase or decrease the importance of a characteristic under others. Comparatively, during the tests made with 151 images and 4530 execution samples for each image in each model, the proposed approach obtained an error rate of 0,49% in the pixel classification versus 3,14% for the control method, without the use of adjustments, in addition to proving itself more stable compared to this performance regarding the different images that were tested and still it obtained some improvement in the model’s final time of processing.
Image segmentation is the process of identifying and separating relevant structures and objects in an image, it is not a trivial task for a computational algorithm due to the complexity of the elements involved in the process. The variety of images and its vast number of characteristics make it impossible to use only a model for the segmentation task. In this way, the application of machine learning algorithms is highlighted, because this branch of artificial intelligence tries to reproduce the way of human learning. The application of the competition and cooperation model among particles in the image segmentations is considered an interesting approach since it has low computational cost, besides reducing the need of specialists for generating labeled data. This work proposes improvements in the making of the complex network that the competition and cooperation model among particles uses, through the alteration of the characteristics extracted from pixels, changes in the shape of the model’s connection of the vertices using the similarity of the characteristics and the image pixel’s spatial neighboring and the influence generated by a particle under the other neighbor vertices within the complex network. The proposed methodology resulted in an improvement of the process’s automation level, eliminating the need to increase or decrease the importance of a characteristic under others. Comparatively, during the tests made with 151 images and 4530 execution samples for each image in each model, the proposed approach obtained an error rate of 0,49% in the pixel classification versus 3,14% for the control method, without the use of adjustments, in addition to proving itself more stable compared to this performance regarding the different images that were tested and still it obtained some improvement in the model’s final time of processing.
Descrição
Palavras-chave
Segmentação de imagens, Aprendizado de máquina, Aprendizado semi-supervisionado, Competição e cooperação entre partículas, Image segmentation, Machine learning, Semi-supervised learning, Particle competition and cooperation