This paper introduced a novel approach for image retrieval, the scene-aware object saliency ranking algorithm(SASR), which addressed the issue of traditional image retrieval techniques relying on semantic similarity and neglecting the crucial importance of object relationships within a scene. SASR consisted of two stages. In the first stage, this paper proposed a viewpoint data-based method called the "combined threshold" to annotate true value labels for object-level saliency ranking, simplifying the annotation of ranking labels. In the second stage, this paper presented an object-level saliency ranking network based on graph convolutional networks that resolved several specific difficulties encountered in sorting objects. The proposed algorithm improved on the current saliency ranking label generation methods and was tested via a large number of comparative experiments. The experimental results on the SALICON dataset show that the SASR algorithm enhances saliency ranking performance significantly. Moreover, the results from the NUS-WIDE dataset indicate that, when supported by the proposed algorithm, image retrieval performance increases by an average of 2%, which solidifies the efficacy of the proposed algorithm.

[2023-04-25] Accepted Paper /doc/2023.01.0028/v1

[2023-10-05] Printed Article /doc/2023.01.0028/v2