Using Knowledge Graph and KD-Tree Random Forest for Image Retrieval

Semantic-based image retrieval has recently become popular and contains many challenges. This work proposes a framework for semantic-based image retrieval by using a Knowledge Graph and a KD-Tree Random Forest. In the preprocessing phase, firstly, we use the R-CNN network to segment the image set into object images, extract features for object images, and classify object images by using the KD-Tree. Secondly, a KD-Tree Random Forest is used to classify the visual relationship of each pair of objects. Thirdly, a Relationship Graph is created based on object images and their visual relationships. Finally, a Knowledge Graph is built and applied to a semantic-based image retrieval system. After segmenting the objects using the R-CNN network for each query image, the KD-Tree is used to classify the objects, and a KD-Tree Random Forest is used to predict their visual relationships. Then, a SPARQL query is generated to retrieve based on a Knowledge Graph and extract similar images. The experiment results with the Oxford Flower-17 and Flickr image sets showed that the precision achieved scores of 0.8586 and 0.8189, respectively. These results are compared with other works to demonstrate the correctness of the proposed method.