Image blurring and sharpening inspired three-way clustering approach

Three-way clustering is a new type of clustering algorithm that divides the clustering results into three different parts or regions. This division allows a clear distinction between the central core and the outer sparse or fringe regions of a cluster. This algorithm is useful in situations when clusters have an unclear and unsharp boundary. In existing studies, a pair of thresholds are typically used to define the three regions of three-way clustering which demands the determination of suitable threshold values. In this paper, we propose an approach called blurring and sharpening based three-way clustering (BS3WC) which constructs the three-way clusters without the need for determining the thresholds. The BS3WC is motivated by observing that the blurring and sharpening operations can produce a three-way representation for a typical object in an image consisting of a core inner, outer blurry, and part not belonging to the object. The BS3WC works in two steps. In step one, it converts a hard cluster into an image. It next defines cluster blur and cluster sharp operations, which are used to create three-way representation for clusters. The BS3WC is validated with 31 datasets including both synthetic and real-life datasets using typical benchmarks of ACC, ARI, NMI and compared with the existing three-way as well as other notable approaches. We also consider the performance of the BS3WC approach in the application area of open-world classification for identifying unknown instances. Experimental results suggest that BS3WC may effectively cluster the data and provide results that are comparable to well-known approaches in the considered application area.

Automated summary

Three-way clustering is a novel clustering algorithm that divides clustering results into core, blurry outer, and non-object areas, suitable for situations with unclear boundaries. Its characteristic of constructing clusters without the need for determining thresholds has been proven to exhibit superior performance in data clustering and open-world classification applications.