A short-term learning framework based on relevance feedback for content-based image retrieval

In this paper a short-term learning method based on relevance feedback for content-based image retrieval is proposed. In content based image retrieval systems, a set of low level features is used to find similar images to the query image. However, the extracted features are not able to represent the content of the images precisely. To resolve this issue, a learning method based on relevance feedback is proposed. The proposed method is built on a short-term learning method which is based on near strangers or distant relatives model. Based on this model, if two individuals are relatives, it is more likely that they have the same characteristics. In our proposed method, after the first iteration of retrieval, in each step of the proposed relevance feedback method, some retrieved images are labeled manually to be related to the query image. Then, new similar images are retrieved based on the labeled images. In the next step, based on the distance of these similar images from the query image, the more similar images to the query image are considered to be the retrieved images for the next iteration. Finally, over iterations, the more similar images to the query are retrieved. The proposed method has been evaluated on Corel-10k dataset which has 10,000 images in 100 different classes. Experimental results show that the precision of the proposed method is significantly higher than the precision of some recently developed methods.