Low-level feature image retrieval using representative images from minimum spanning tree clustering

Piyavach Khunsongkiet, Jakramate Bootkrajang, Churee Techawut

Abstract: Typical content-based image retrieval systems retrieve images based on comparison of low-level features such as images color, texture, and shapes of objects in the images. Further, the image covariance descriptor (CD) and the image Patch Relational Covariance Descriptor (PRCD) can be used to summarize low–level features and the visual arrangement to improve the precision of the retrieval. Nonetheless, comparing images based on those two descriptors is computationally expensive. Therefore, this research proposes a clustering method that dynamically groups database images using the Minimum Spanning Tree Clustering algorithm (MSTC). The technique is named Representative Images from Minimum Spanning Tree Clustering (RIMSTC). In the proposed technique, only the representative images selected from each cluster are compared with the input image . Experimental results demonstrated that the proposed representative images by COV and PRCD combined with RIMSTC helps to improve the retrieval time while maintaining comparable retrieval performance to existing methods.