一种有效的层次分割方法及其在物体定位中的应用

An Efficient Hierarchical Image Segmentation Approach and its Application on Object Localization

为了提高图像层次分割方法的通用性和计算效率,提出一种有效的层次分割方法,并探索了其在物体定位中的应用.在经典的g Pb-owt-ucm分割框架下,利用非最大值抑制的边缘改进分水岭变换算法,并采用高性能的边缘检测算子构建了完整的层次分割方法;然后将该分割方法应用在一种物体定位方法中,通过边缘检测尺度和超度量轮廓图分割门限上的变化重新设计了多样性策略,并提出了一种基于分类器的物体假设排序方法.在BSDS500分割数据集上的实验结果表明,文中方法能以更快的速度生成高质量的图像分割,同时无需边缘方向信号;在Pascal VOC 2007数据集上的实验结果表明,改进后的物体定位方法能产生更精确的物体假设区域.

To improve the generality and computational efficiency of hierarchical image segmentation method, this paper proposes an efficient hierarchical segmentation approach, and explores its application on object localization. In the g Pb-owt-ucm framework, we improve the watershed transform algorithm using non-maximal suppressed edge, and then construct the full version of hierarchical segmentation with top-performing edge detector. Then, the hierarchical segmentation approach is used for object localization algorithm Selective Search. The diversity strategy of Selective Search is redesigned by changing the edge detection scale and UCM threshold, and an object proposals ranking method based on the classifier is proposed. The experiment on BSDS500 shows that our approach generates the high quality segmentation in a higher speed without the need of edge direction signal. The experiment on Pascal VOC 2007 shows that the improved object localization algorithm can generate more precise object proposals.