A density clustering approach for CBIR system

Abstract

Searching an image in a huge set of images became an important task in several domains such as crime, medicine, geology and so on. The task of retrieving images by their visual contents is called content-based image retrieval (CBIR) systems. These systems have to be fast, efficient and semantically similar. For this aim, we used a new density clustering technique in our proposed CBIR system. The paper describes a new CBIR that uses a t-SNE (t-Distributed Stochastic Neighbor Embedding) data reduction and a proposed density-based clustering method. Several advantages are deduced from the proposition. First, reducing the dimensionality minimizes the required time and storage space. Next, reducing images to a very low dimension such as 2D or 3D permits an easier visualization. Also, no need to set image data parameters for clustering. Likewise, No need to introduce the number of clusters. Besides, it is effective for several image data especially shaped data. For validation tests, we use ZUBUD, Wang databases and shape datasets. Several comparison with two other CBIR systems such as FIRE and LIRE are included. The results obtained demonstrate the originality, reliability, and relevance of our proposition.