Using neural network and k-means cluster for image segmentation in outdoor scenes

Resumo

 The present paper discusses the image processing algorithms that treat the problem of image segmentation. The idea is evaluate segmentation algorithms performance applied to captured images in outdoor scenes under natural illumination. This light source type can cause low image quality. Both, K-means cluster and neural network based on pixel RGB space colour are described like segmentation algorithms. The neural network and K-means performance evaluation will be measured by a method based on discrepancy. In order to handle digital images, a computational environment is then implemented. Besides the image capture and the pre-processing algorithms, this environment is composed by the K-means, neural network and discrepancy method applied on performance evaluation, which is the main theme of this paper. The digital images describe plants species and land captured by a video camera (CCD - Charged Coupled Device). These images will then be segmented with the purpose of separating the plant species from the land regions. With the K-means algorithm, one may get some faulty segmented pixels while the use of neural network techniques may increase the performance but some errors might remain. The paper introduces the combination of both these methods and shows that a much better performance can be achieved.