Hierarchical Salient Point Selection for image retrieval

Salient point detection in images is very useful for image processing applications like image compression, object detection and object recognition. It is also frequently used to represent local properties of the image in content-based image retrieval (CBIR). Many research results focused on finding the most salient points in the image. However, the large number of salient points and continuous point sets are still the problems. Based on the saliency values from wavelet-based methods, this paper presents a hierarchical algorithm for selecting the most salient points such that they cannot only give a satisfying representation of an image, but also make the image retrieval systems more efficiently. Under a top–down approach from quadtree data structure, the algorithm keeps the most salient points in each quadrant according to the percentage of saliency values in the whole image. The performance of the proposed method was evaluated with the spreading measure and retrieval rate from a CBIR system. In this experiment, it shows that our method is robust and the extracted salient points provide efficient retrieval performance comparing with two wavelet-based point detectors.

► Salient points in images are very useful for content-based image retrieval. ► Large number of salient points and continuous point sets are still the problems. ► A top–down hierarchical approach is proposed to find the most salient points. ► The hierarchical algorithm can be applied to other salient point detectors.