Volume data classification is a core issue of transfer function in volume rendering. Scalar-gradient magnitude histogram of volume is a classic feature space, and has been applied in volume classification for its nice result in visual extraction of boundaries between different materials. However, the design of transfer function based on scalar-gradient histogram has proven as a time-consuming and complex task which is hard for users to conduct interactions. In this paper, scalar-gradient histogram is treated as a density distribution of all voxels. This approach assumes that the density of a material center is higher than their neighbors and the distance between two material centers is far enough. By computing the minimum distance between each points and all other points with higher density in scalar-gradient histogram, a density-distance graph is constructed based on densities and minimum distances of all points. The density peaks are easily observed in the graph and can guide the users to select centers of each material as a progressive volume classification process through a set of specified interactions. Experimental results demonstrate that the presented approach does not require the prior knowledge of categories, and the volume classification is accurate with high performance.