特征空间中的边缘检测及样本选择

将子空间分类法拓展到特征空间后，与核主成分分析结合提出了一种边缘检测的方法及其训练样本选择策略。是基于特征空间中的核方法，对图像特征表达建立了统一的模型，可处理非高斯分布的数据。可与经典的边缘检测算子或其他方法相结合，增强边缘检测的效果和稳定性。只需训练一次，便可将边缘特征从一幅与训练图完全不同的测试图中提取出来。实验结果表明，对噪声有很好的鲁棒性，能很好地适应小样本训练，其边缘检测的效果明显比经典算子，主成分分析，非线性主成分分析的效果好。

Edge Detection Approaches in Feature Space and Strategy for Samples Selecting

To build a consistent model to represent image feature using kernel method in the feature space, an edge detection approach based on kernel principal component analysis (KPCA) and the subspace classification in the feature space was proposed, it can treat the Non-Gaussian distribution data. The KPCA was applied to train classifiers for edge detection, and the subspace classification in feature space was proposed to extract the edge feature from a testing image which was completely different from the training one. Then a strategy for sample selecting was proposed to stabilize the effect of our edge detection approach by combining the approach with the classical operators. The simulating experiment showed that the approach proposed in this paper had a better performance than nonlinear PCA, especially when the number of samples was very small. And the approach was robust to noise.