Unsupervised evaluation of segmentation quality is a crucial step in imagesegmentation applications. Previous unsupervised evaluation methods usuallylacked the adaptability to multi-scale segmentation. A scale-constrainedevaluation method that evaluates segmentation quality according to thespecified target scale is proposed in this paper. First, regional saliency andmerging cost are employed to describe intra-region homogeneity and inter-regionheterogeneity, respectively. Subsequently, both of them are standardized intoequivalent spectral distances of a predefined region. Finally, by analyzing therelationship between image characteristics and segmentation quality, weestablish the evaluation model. Experimental results show that the proposedmethod outperforms four commonly used unsupervised methods in multi-scaleevaluation tasks.
展开▼
