Contour detection based on support vector regression in natural images

Contour detection in natural images has been studied for years and is still a difficult problem, because the relationship between the actual contour and the contour signals (such as gradient) extracted from various image channels is confusing. Some recent works attempt to learn a linear regression model to combine the available contour signals into a single measure, which often lead to an undesirable solution if given nonlinear signals. In this paper, we propose a new contour detection method based on Support Vector Regression (SVR). While training, the SVR with a nonlinear kernel function is used to combine the various contour signals together nonlinearly. While predicting, our method outputs an integrated contour signal directly and then the non-maximum suppression is employed to generate the thin contour. We learned and tested on Berkeley Segmentation Dataset and Benchmark 500 database (BSDS500). The experimental results show that the new method outperforms the state-of-the-arts method with regard to precision-recall curve and F-measure.