Multi-object image retrieval based on R-CNN network and KD-Tree structure

Abstract

Multi-object image retrieval is a crucial problem in the field of image retrieval because of the diversity and complexity of digital images. In this paper, a method of multi-object image retrieval based on the R-CNN network with a KD-Tree structure is proposed to benefit from the advantages of the R-CNN network in identifying and classifying each object on the image separately; at the same time, the KD-Tree structure has high storage capacity and stable retrieval time. To solve this problem, we extracted the objects on the image data set, classified them with the R-CNN network model, and stored them on the KD-Tree structure. Then, each input image was segmented according to each object; the feature vector was extracted, and a similar image set was retrieved based on the KD-Tree structure. On this basis, a model of image retrieval using the R-CNN network and KD-Tree structure was proposed. To demonstrate the correctness of the proposed theoretical basis, we developed an experiment on the COCO image data set with an image retrieval precision of 0.6898. The experimental results were compared with other works on the same data set. This comparison proves the feasibility and effectiveness of the proposed method, which can be applied to multi-object images.