Enhance Neighbor Reversibility in Subspace Learning for Image Retrieval

Two images that describe similar content usually have the neighbor-reversibility (NR) correlation, i.e., each image is among the neighbors of the other one. This phenomenon can be frequently observed in image retrieval. Some previous works have successfully utilized the NR correlation to improve search accuracy. In these methods, the retrieved images that have the NR correlation with the query image will rank ahead in the returned list. In this manner, the chance of finding the positive result in the top position will be increased. However, these methods will suffer from high computational cost when they deal with massive images. During the training phase, these methods need to calculate a weighting factor for each database image. When a database grows to be huge, this calculation will be disastrous. In order to overcome this shortcoming, we want to design a new method that can enhance the NR correlation without utilizing the whole database images for training. Our idea is to enhance the NR correlation within subspace learning. In the proposed method, we utilize a small number of training images to learn a low-dimensional subspace, where the NR correlation is best preserved. Then, images are transformed from their global features into the low-dimensional representation with the learned mapping function. In this manner, the efficiency of training will be improved. Experiment on four public retrieval datasets demonstrates that our method compares favorably with the subspace learning baseline methods for image retrieval.